# Hugging Face

The Hugging Face blog

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost

DevFeed: [Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost](<https://devfeed.tech/articles/introducing-harness-adapters-adapt-your-favorite-model-to-your-harness-to-improve-performance-and-cost-146558.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/harness-adapters-at-scale>)

Author: Younes B; ben burtenshaw; Nathan Habib; Lysandre; Amine Dirhoussi; Adithya S K; shaun smith; Sergio Paniego; Joel Niklaus; Lewis Tunstall

Published: 2026-11-20T00:00:00Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [adapters](<https://devfeed.tech/tags/adapters.md>), [agent-harness](<https://devfeed.tech/tags/agent-harness.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [agents](<https://devfeed.tech/tags/agents.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [cost](<https://devfeed.tech/tags/cost.md>), [harness](<https://devfeed.tech/tags/harness.md>), [lora](<https://devfeed.tech/tags/lora.md>), [model](<https://devfeed.tech/tags/model.md>), [open](<https://devfeed.tech/tags/open.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [openenv](<https://devfeed.tech/tags/openenv.md>), [performance](<https://devfeed.tech/tags/performance.md>), [rl](<https://devfeed.tech/tags/rl.md>), [source](<https://devfeed.tech/tags/source.md>), [trl](<https://devfeed.tech/tags/trl.md>)

### AI overview

The article reports that reinforcement learning inside each target agent harness improved Qwen3.8-27B's coding benchmark results and reduced tool use. Separate LoRA adapters trained for Pi, Hermes Agent, and OpenCode each beat the base model under their own harness on Terminal-Bench 2.1 and hard SWE-bench Pro tasks. On tasks both models solved, specialists used 53-72% fewer tool calls. A merged adapter also beat the base under all three harnesses; with thinking off it averaged 39.5% to 49.0% pass@1, while cutting tool calls by 53-73%. The recipe uses validated Harbor tasks, difficulty filtering so GRPO receives tasks the base model solves inconsistently, and TRL's AsyncGRPOTrainer with the harness running as shipped. Results are not directly comparable across harnesses because their defaults differ. On 300 AutomationBench tasks outside the training domain and harnesses, adapters stayed near the base model, so transfer to that setting was not clear. The authors also note possible reasoning leakage in non-thinking outputs and call for further study.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

DevFeed: [DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency](<https://devfeed.tech/articles/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency-143008.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin>)

Author: Jason Chen

Published: 2026-11-18T19:54:56Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [post-training](<https://devfeed.tech/topics/post-training.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [distributed-training](<https://devfeed.tech/topics/distributed-training.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Inference](<https://devfeed.tech/topics/inference.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-tasks](<https://devfeed.tech/tags/agent-tasks.md>), [apache-2-0](<https://devfeed.tech/tags/apache-2-0.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [backend](<https://devfeed.tech/tags/backend.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [coding](<https://devfeed.tech/tags/coding.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [qwen3-8](<https://devfeed.tech/tags/qwen3-8.md>), [qwen3-8-27b](<https://devfeed.tech/tags/qwen3-8-27b.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [sessions](<https://devfeed.tech/tags/sessions.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [terminal](<https://devfeed.tech/tags/terminal.md>), [terminal-bench](<https://devfeed.tech/tags/terminal-bench.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

RadixArk reports that DaisyRose-27B, built on Qwen3.8-27B and trained with Miles, raised Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. Training used 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs; the tasks targeted original-model errors, required multi-step work, and were checked to ensure successful completion was necessary. The model retains the base architecture, tokenizer, chat template, image understanding, and 256K context window; checkpoints are Apache 2.0. RadixArk says its internal backend achieved 37% MFU versus 16% for the open-source Megatron-based trainer on comparable runs, and roughly twice the end-to-end throughput. Its sandbox service reported 99% creation success over nine weeks, exceeding 99.95% since late October, and about 80% lower infrastructure costs than popular cloud sandbox providers. The article recommends SGLang with DSpark speculative decoding and points to the model card for inference settings and evaluation details.

### Source excerpt

Today we are releasing **DaisyRose-27B**, RadixArk's first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles. We trained it on long-horizon agent tasks with sessions up to 1M tokens, with just 1,200 curated tasks in 18 days on 144 H200 GPUs. DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from **5.6% to 12.0%**. Beyond Terminal-Bench 4.1, it also delivers gains on a variety of other reasoning, instruction following, and coding benchmarks.

## Time-Series Models in the Streamhouse: A New Frontier and Use Cases

DevFeed: [Time-Series Models in the Streamhouse: A New Frontier and Use Cases](<https://devfeed.tech/articles/time-series-models-in-the-streamhouse-a-new-frontier-and-use-cases-143009.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ibm-research/time-series-powered-streamhouse>)

Author: Nicholas Fuller; Sean Falconer; Ayhan Sebin; Pankaj Dayama; Roman Vaculin; Bilge Zeren Aksu; Julian Payne; Mountu Jinwala; Wesley M Gifford; Chandra Reddy; Vijay Ekambaram; Arindam Jati; Robert Metzge

Published: 2026-11-18T15:00:23Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Time Series](<https://devfeed.tech/topics/time-series.md>), [streaming data analytics](<https://devfeed.tech/topics/streaming-data-analytics.md>), [Event-Streaming](<https://devfeed.tech/topics/event-streaming.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Confluent Platform](<https://devfeed.tech/topics/confluent-platform.md>)

Tags: [anomaly-detection](<https://devfeed.tech/tags/anomaly-detection.md>), [business](<https://devfeed.tech/tags/business.md>), [can](<https://devfeed.tech/tags/can.md>), [cases](<https://devfeed.tech/tags/cases.md>), [change](<https://devfeed.tech/tags/change.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [confluent](<https://devfeed.tech/tags/confluent.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [cpus](<https://devfeed.tech/tags/cpus.md>), [detection](<https://devfeed.tech/tags/detection.md>), [downloads](<https://devfeed.tech/tags/downloads.md>), [forecasting](<https://devfeed.tech/tags/forecasting.md>), [form](<https://devfeed.tech/tags/form.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [help](<https://devfeed.tech/tags/help.md>), [models](<https://devfeed.tech/tags/models.md>), [series](<https://devfeed.tech/tags/series.md>), [time](<https://devfeed.tech/tags/time.md>), [time-series](<https://devfeed.tech/tags/time-series.md>), [use-cases](<https://devfeed.tech/tags/use-cases.md>)

### AI overview

IBM time-series foundation models are available as native Confluent functions in managed Confluent Cloud and on-premises Confluent Platform for Apache Flink, running on CPUs. Flink can apply forecasting, anomaly detection, and time-series representations to streaming data with SQL, putting results alongside current events. Pretrained models offer a starting point across many series; workflows can incorporate recent observations, business rules, and human review. Examples include forecasting production quality to evaluate setpoints, flagging unusual store sales against a chosen confidence level, refreshing store-level demand forecasts to inform reorder points, adding individual spending patterns to existing payment checks, and pacing trades from short-term volume forecasts. Similarity search with TSPulse requires a separately built embedding pipeline and index, and can help investigate patterns but cannot establish root cause. These scenarios are illustrative; outcomes depend on model requirements, pipeline configuration, and application logic.

### Source excerpt

**Real-time forecasting, anomaly detection, and optimization, powered by IBM time-series models with more than 50 million downloads. Now available natively on Confluent, in the cloud and on premises, running on CPUs.** Every business generates a continuous record of change: a machine reading, a payment, a sale, an order, or a change in inventory. These observations form time series, and they carry signals that can help teams anticipate what happens next.

## 180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers

DevFeed: [180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers](<https://devfeed.tech/articles/180b-parameters-32gb-of-ram-pushing-moe-inference-to-the-limit-with-transformers-143010.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/moe-expert-streaming>)

Author: Tarek Ziadé

Published: 2026-11-18T00:00:00Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Local AI](<https://devfeed.tech/topics/local-ai.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>), [On-device AI](<https://devfeed.tech/topics/on-device-ai.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [quantization](<https://devfeed.tech/topics/quantization.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [apple](<https://devfeed.tech/tags/apple.md>), [apple-silicon](<https://devfeed.tech/tags/apple-silicon.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [cache](<https://devfeed.tech/tags/cache.md>), [inference](<https://devfeed.tech/tags/inference.md>), [moe](<https://devfeed.tech/tags/moe.md>), [ram](<https://devfeed.tech/tags/ram.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

The excerpt reports a Transformers prototype that streams only router-selected MoE experts from SSD through a bounded device cache, remaps expert IDs to shared cache slots, and reuses `grouped_mm` for expert computation. On a 32 GB M5, after quantizing experts to 3 bits, compressing dense layers, and adding a separate disk path for Qwen3.8-Flash-Next's n-gram table, the system reached about 14 tokens/s--roughly 6% behind Strata on short decode, with slightly faster prefill. It uses about 18 GiB after loading; a 6K prompt peaks at 25.4 GiB, and the author reports needing to close the browser to avoid swapping. The remaining gap is attributed mainly to host round trips and GPU idle time between layers; closing it would require deeper MPS synchronization control unavailable in PyTorch. The API is still under discussion, and the approach currently supports models with the standard experts layout.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Hugging Face and OpenRouter join forces to make open-weight models easy to run

DevFeed: [Hugging Face and OpenRouter join forces to make open-weight models easy to run](<https://devfeed.tech/articles/hugging-face-and-openrouter-join-forces-to-make-open-weight-models-easy-to-run-137049.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/openrouter>)

Author: Célina Hanouti; Victor Mustar; Bertrand Chevrier; Lucain Pouget; Pierric Cistac; Eliott Coyac; Julien Chaumond

Published: 2026-11-16T00:00:00Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [inference-providers](<https://devfeed.tech/topics/inference-providers.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>), [litellm](<https://devfeed.tech/topics/litellm.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [agent](<https://devfeed.tech/tags/agent.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [announcement](<https://devfeed.tech/tags/announcement.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [apis](<https://devfeed.tech/tags/apis.md>), [artificial-analysis](<https://devfeed.tech/tags/artificial-analysis.md>), [artificial-analysis-intelligence-index](<https://devfeed.tech/tags/artificial-analysis-intelligence-index.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [audio](<https://devfeed.tech/tags/audio.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [best](<https://devfeed.tech/tags/best.md>), [hub](<https://devfeed.tech/tags/hub.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [inference-providers](<https://devfeed.tech/tags/inference-providers.md>), [intelligence](<https://devfeed.tech/tags/intelligence.md>), [join](<https://devfeed.tech/tags/join.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [partnerships](<https://devfeed.tech/tags/partnerships.md>), [science](<https://devfeed.tech/tags/science.md>)

### AI overview

Hugging Face now routes language-model requests through OpenRouter, which selects a provider using live price, speed, and reliability measurements and retries another provider when a call fails. Existing code and Hugging Face tokens continue to work; supported interfaces include OpenAI Chat Completions and Responses and Anthropic Messages. Users can select a routing policy or provider with a model-ID suffix, disable providers in settings, and use `X-HF-Session-Id` to keep a conversation on the provider holding its prompt cache. Pricing remains the provider's price with no Hugging Face markup, but users can no longer use their own provider API keys for language models. Image, video, and audio generation remain directly on Hugging Face Inference Providers. OpenRouter reports that its rebuilt routing cut capacity errors by 46% and improved time to first token by 16.9% and throughput by 16.3%. Open-weight models average about four times less per token in the cited figures, though the best open-weight model still scores below the best proprietary model on the cited benchmark, at substantially lower cost per task.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Build a Robot Pick-and-Place Simulation with SimReady Assets

DevFeed: [Build a Robot Pick-and-Place Simulation with SimReady Assets](<https://devfeed.tech/articles/build-a-robot-pick-and-place-simulation-with-simready-assets-133556.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/nvidia/simready-assets-robot-pick-and-place>)

Author: Dillon Bailey; Pomi Lee; Rishabh Chadha

Published: 2026-11-13T18:00:39Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Isaac Sim](<https://devfeed.tech/topics/isaac-sim.md>), [Robotics Simulation](<https://devfeed.tech/topics/robotics-simulation.md>), [Robot Manipulation](<https://devfeed.tech/topics/robot-manipulation.md>), [Omniverse](<https://devfeed.tech/topics/omniverse.md>), [Isaac](<https://devfeed.tech/topics/isaac.md>), [Agent Skills](<https://devfeed.tech/topics/agent-skills.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [Tutorial](<https://devfeed.tech/topics/tutorial.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [arm](<https://devfeed.tech/tags/arm.md>), [asset](<https://devfeed.tech/tags/asset.md>), [assets](<https://devfeed.tech/tags/assets.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [build](<https://devfeed.tech/tags/build.md>), [can](<https://devfeed.tech/tags/can.md>), [compose](<https://devfeed.tech/tags/compose.md>), [controls](<https://devfeed.tech/tags/controls.md>), [example](<https://devfeed.tech/tags/example.md>), [isaac-sim](<https://devfeed.tech/tags/isaac-sim.md>), [openusd](<https://devfeed.tech/tags/openusd.md>), [pick-and-place](<https://devfeed.tech/tags/pick-and-place.md>), [properties](<https://devfeed.tech/tags/properties.md>), [robot](<https://devfeed.tech/tags/robot.md>), [robotics](<https://devfeed.tech/tags/robotics.md>), [run](<https://devfeed.tech/tags/run.md>), [simulation](<https://devfeed.tech/tags/simulation.md>), [use-case](<https://devfeed.tech/tags/use-case.md>), [uses](<https://devfeed.tech/tags/uses.md>)

### AI overview

The walkthrough uses SimReady Central assets to build and test a scripted beverage-can pick-and-place task in Isaac Sim with a UR17G arm and Robotiq 2F-140 gripper. SimReady profiles define use-case features and requirements; validation checks authored settings such as joint drives, while the assembled simulation tests whether the robot can grasp, move, and release the can. The workflow downloads complete asset folders, assembles and validates the arm and gripper, then runs agent-generated controls, checking joint settings, motion, contact, and whether the can stays held. Isaac Sim 6.2 and an agent able to use its skills are required for the described agent-assisted setup; Hugging Face CLI is optional for command-line downloads. The example uses a fixed setup, scripted waypoints, and a 2 N-m drive-effort cap selected through simulation tests. It omits vision, limits collision checks, and does not model can deformation or liquid motion. Results do not establish robustness to uncertain positions, other payloads, or real hardware.

### Source excerpt

This walkthrough showcases a beverage can pick-and-place task that uses the UR17G arm, Robotiq 2F-140 gripper, and beverage can to compose a scene in Isaac Sim, check their behavior, and run a pick-and-place example with agent-generated controls. For robotics workflows, SimReady means an OpenUSD asset has been prepared with the properties needed for a defined simulation use case, such as collisions, mass, joints, and joint drives.

## XYBench: Helping with the problem behind the question

DevFeed: [XYBench: Helping with the problem behind the question](<https://devfeed.tech/articles/xybench-helping-with-the-problem-behind-the-question-133554.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/allenai/xybench>)

Author: Kyle Wiggers

Published: 2026-11-13T17:02:35Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [gaia](<https://devfeed.tech/topics/gaia.md>), [Benchmark](<https://devfeed.tech/topics/benchmark.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-assistants](<https://devfeed.tech/tags/ai-assistants.md>), [ai-systems](<https://devfeed.tech/tags/ai-systems.md>), [award](<https://devfeed.tech/tags/award.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [carnegie-mellon](<https://devfeed.tech/tags/carnegie-mellon.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-sonnet](<https://devfeed.tech/tags/claude-sonnet.md>), [cooking](<https://devfeed.tech/tags/cooking.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [emnlp-2026](<https://devfeed.tech/tags/emnlp-2026.md>), [english](<https://devfeed.tech/tags/english.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [gemini](<https://devfeed.tech/tags/gemini.md>), [gemini-3-1-pro](<https://devfeed.tech/tags/gemini-3-1-pro.md>), [misconceptions](<https://devfeed.tech/tags/misconceptions.md>), [xybench](<https://devfeed.tech/tags/xybench.md>)

### AI overview

XYBench is a benchmark of 8,115 real-world and generated questions designed to test whether AI assistants spot mistaken assumptions, address users' underlying goals, and make goal-focused solutions central to their answers. It evaluates solution coverage, answer emphasis, and misconception identification. In single-turn tests, five models more often addressed the literal request than the inferred goal; on Stack Overflow, they focused primarily on the underlying goal in 35-41% of responses, versus 47-57% focused on the stated request. Giving Sonnet 4.6 the intended goal raised underlying-goal solution coverage from 59% to 67%, but lowered misconception identification from 47% to 34%. In a separate choice task, six models selected goal-focused answers 71-84% of the time, revealing a gap between recognizing and producing such advice. The benchmark covers English queries, mainly procedural problems; broader languages, cultures, and conversational settings remain open directions.

### Source excerpt

Truly useful AI assistants need to be able to help users who ask misguided questions--not just users who ask the right question with the right amount of detail. We're releasing **XYBench**, a benchmark of 8K+ questions that tests how AI assistants respond to requests based on misconceptions, such as asking how to thicken frosting that keeps sliding off cupcakes when the real problem is that the cupcakes are still too warm to frost. XYBench won the Best Resource Paper Award at EMNLP 2026.

## llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem

DevFeed: [llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem](<https://devfeed.tech/articles/llama-cpp-v1-0-0-a-stronger-foundation-for-the-local-ai-ecosystem-133555.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ggml-org/llama-cpp-v1>)

Author: Daniel Bevenius; Georgi Gerganov; Merve

Published: 2026-11-13T14:05:22Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [ggml](<https://devfeed.tech/topics/ggml.md>)

Tags: [accessible](<https://devfeed.tech/tags/accessible.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [ai-products](<https://devfeed.tech/tags/ai-products.md>), [backends](<https://devfeed.tech/tags/backends.md>), [building](<https://devfeed.tech/tags/building.md>), [c](<https://devfeed.tech/tags/c.md>), [c-c-plus-plus](<https://devfeed.tech/tags/c-c-plus-plus.md>), [c-plus-plus](<https://devfeed.tech/tags/c-plus-plus.md>), [consumer](<https://devfeed.tech/tags/consumer.md>), [cross-platform](<https://devfeed.tech/tags/cross-platform.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [implementation](<https://devfeed.tech/tags/implementation.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llama](<https://devfeed.tech/tags/llama.md>), [llama-cpp](<https://devfeed.tech/tags/llama-cpp.md>), [local](<https://devfeed.tech/tags/local.md>), [local-ai](<https://devfeed.tech/tags/local-ai.md>), [v1](<https://devfeed.tech/tags/v1.md>)

### AI overview

llama.cpp v1.0.0 marks a commitment to clearer compatibility boundaries for downstream developers: future API and ABI incompatibilities in llama.cpp and ggml will be reflected in major-version bumps. The project describes llama.cpp as a portable C/C++ inference platform using quantization and optimized backends across a wide range of hardware. It reports support for more than 150 model architecture variants and mainstream speculative decoding methods, with ggml backends that can be compiled as dynamic libraries and optionally loaded at runtime. The excerpt also describes expanded CI infrastructure and hardware runners to check changes, while emphasizing continued community contributions and support for model architectures.

### Source excerpt

llama.cpp makes it possible to run multi-modal LLMs across a wide range of hardware - from consumer-grade local machines to high-end enterprise servers. Through quantization, optimized backends, and a portable C/C++ implementation, we make AI inference accessible at every level of the software stack. We believe this is the most lightweight, versatile and sensible platform for building cross-platform and private AI products.

## LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

DevFeed: [LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B](<https://devfeed.tech/articles/llmon-an-llm-native-markup-language-now-in-granite-4-1-3b-and-8b-131215.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite>)

Author: Abraham Daniels; Basel Shbita; Chad DeLuca; Dan Gutfreund; Michael Hind; Shub A

Published: 2026-11-12T19:07:35Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [markup](<https://devfeed.tech/topics/markup.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>), [AI Prompt Templates](<https://devfeed.tech/topics/ai-prompt-templates.md>), [JSON](<https://devfeed.tech/topics/json.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [dpo](<https://devfeed.tech/topics/dpo.md>)

Tags: [act](<https://devfeed.tech/tags/act.md>), [allowed](<https://devfeed.tech/tags/allowed.md>), [components](<https://devfeed.tech/tags/components.md>), [context](<https://devfeed.tech/tags/context.md>), [data](<https://devfeed.tech/tags/data.md>), [either](<https://devfeed.tech/tags/either.md>), [formats](<https://devfeed.tech/tags/formats.md>), [generate](<https://devfeed.tech/tags/generate.md>), [granite](<https://devfeed.tech/tags/granite.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [json](<https://devfeed.tech/tags/json.md>), [language](<https://devfeed.tech/tags/language.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llms](<https://devfeed.tech/tags/llms.md>), [markup](<https://devfeed.tech/tags/markup.md>), [model](<https://devfeed.tech/tags/model.md>), [native](<https://devfeed.tech/tags/native.md>), [next](<https://devfeed.tech/tags/next.md>), [over](<https://devfeed.tech/tags/over.md>), [prompt](<https://devfeed.tech/tags/prompt.md>)

### AI overview

LLMON uses reserved special-token delimiters to mark context spans such as instructions, data, and requirements, then binds selected spans through an execution block. This makes relationships explicit for the model and is intended to reduce confusion and static indirect prompt injection. In the described OPI benchmark, LLMON previews nearly eliminated attacks, avoiding 47% more than the best tested frontier model and 95% more than the best open-source model; the authors report comparable general benchmark performance to Granite 4.1 Instruct. The gains are qualified: the benchmarks cover static injections, adaptive multi-turn attacks remain under study, and the learned privilege rule is not an enforced guarantee. The experimental 3B and 8B previews include a chat template, so naming message parts can be enough to use LLMON; converters also support lossless JSON conversion. The authors say it is for exploration, not production.

### Source excerpt

LLMs rely on their context to generate the next set of tokens. Over a long trajectory, the context fills with numerous types of components: a system prompt, past conversation turns, instructions that act on retrieved documents, tool responses or user data, requirements constraining the allowed output, and reasoning chains. The context carries a rich structure, most of which is either implicit to the model or encoded in formats such as JSON that were not designed to be digested by LLMs.

## The Definitive Fine-Tuning Guide for 3D Generation Models

DevFeed: [The Definitive Fine-Tuning Guide for 3D Generation Models](<https://devfeed.tech/articles/the-definitive-fine-tuning-guide-for-3d-generation-models-131214.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/3d-lora-finetuning>)

Author: Suvaditya Mukherjee; merve; Pedro Cuenca

Published: 2026-11-12T00:00:00Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [lora](<https://devfeed.tech/topics/lora.md>), [peft](<https://devfeed.tech/topics/peft.md>), [3D](<https://devfeed.tech/topics/3d.md>), [AI Development](<https://devfeed.tech/topics/ai-development.md>), [blender](<https://devfeed.tech/topics/blender.md>)

Tags: [3d](<https://devfeed.tech/tags/3d.md>), [3d-gaussian-splatting](<https://devfeed.tech/tags/3d-gaussian-splatting.md>), [3d-model](<https://devfeed.tech/tags/3d-model.md>), [3dgs](<https://devfeed.tech/tags/3dgs.md>), [a100](<https://devfeed.tech/tags/a100.md>), [adapter](<https://devfeed.tech/tags/adapter.md>), [adapters](<https://devfeed.tech/tags/adapters.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [attention](<https://devfeed.tech/tags/attention.md>), [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [background](<https://devfeed.tech/tags/background.md>), [backpropagation](<https://devfeed.tech/tags/backpropagation.md>), [bf16](<https://devfeed.tech/tags/bf16.md>), [blender](<https://devfeed.tech/tags/blender.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [finetuning](<https://devfeed.tech/tags/finetuning.md>), [generation](<https://devfeed.tech/tags/generation.md>), [guide](<https://devfeed.tech/tags/guide.md>), [lora](<https://devfeed.tech/tags/lora.md>), [models](<https://devfeed.tech/tags/models.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [science](<https://devfeed.tech/tags/science.md>)

### AI overview

The guide fine-tunes TRELLIS.2 for footwear geometry with LoRA adapters on its sparse-structure and shape flow DiTs, while freezing the texture DiT because the 188-object GSO subset lacks separate material maps and the task concerns shape. It describes preprocessing meshes into latents and rendering 16 conditioning views per object, then training the two adapters in separate runs. Hugging Face Jobs or local Docker are supported; setup requires an HF_TOKEN, DINOv3 access, and Hub-hosted meshes for a custom dataset. On nine held-out shoes, both adapters improved F-score from 0.276 to 0.803, reduced Chamfer distance from 0.174 to 0.033, and raised normal consistency from 0.530 to 0.804. The authors caution that training loss alone does not establish geometry quality; the reported gains are for this small footwear evaluation.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

DevFeed: [BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model](<https://devfeed.tech/articles/botpsy-on-device-personal-assistant-with-a-sub-billion-parameter-agentic-language-model-129791.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/qvac/botpsy>)

Author: Milan Gritta; Jihye Back; JinUk; Cho; Ohjoon Kwon; Patrik Lambert; Mathias Buus; Amril Nurman

Published: 2026-11-11T05:01:16Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>), [rlvr](<https://devfeed.tech/topics/rlvr.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [online privacy](<https://devfeed.tech/topics/online-privacy.md>), [quantization](<https://devfeed.tech/topics/quantization.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [costs](<https://devfeed.tech/tags/costs.md>), [device](<https://devfeed.tech/tags/device.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [on-device](<https://devfeed.tech/tags/on-device.md>), [privacy](<https://devfeed.tech/tags/privacy.md>)

### AI overview

BotPsy is a 0.8B-parameter, on-device tool-use assistant built on Qwen3.5-0.8B. In the authors' tests against four other models, it roughly doubled its base model's scores on several agent benchmarks and led the tested models on five of six overall; Qwen3.5-2B-Think scored higher on Claw-Eval and some search and special-case tasks. BotPsy also improved instruction following and reduced unsafe responses in automated red-teaming, including under prompt injection, though these results are benchmark-specific. It was trained through supervised fine-tuning, teacher distillation, and two reinforcement-learning stages using tool-use and general-domain tasks. The authors describe it as most suitable for low- to medium-difficulty tasks, primarily in English. Multi-step research remains difficult, and the model makes mistakes; it is not intended for high-stakes or safety-critical decisions or sensitive applications.

### Source excerpt

Frontier cloud LLMs carry high operational costs that can fluctuate unexpectedly, and relying on such third-party models introduces vulnerabilities, as providers can modify, deny, or deprecate access at any time. Furthermore, proprietary models offer limited customization for specific workflows, while posing risks to privacy, data control, and transparency.

## Token Pooling-Aware Training for Multi-Vector Embedding Models

DevFeed: [Token Pooling-Aware Training for Multi-Vector Embedding Models](<https://devfeed.tech/articles/token-pooling-aware-training-for-multi-vector-embedding-models-129792.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/token-pooling>)

Author: Tom Aarsen

Published: 2026-11-11T00:00:00Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [sentence-transformers](<https://devfeed.tech/topics/sentence-transformers.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [Vectors & LLMs](<https://devfeed.tech/topics/vectors-llms.md>)

Tags: [community](<https://devfeed.tech/tags/community.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [embedding-models](<https://devfeed.tech/tags/embedding-models.md>), [embeddings](<https://devfeed.tech/tags/embeddings.md>), [guide](<https://devfeed.tech/tags/guide.md>), [nlp](<https://devfeed.tech/tags/nlp.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [sentence-transformers](<https://devfeed.tech/tags/sentence-transformers.md>), [training](<https://devfeed.tech/tags/training.md>), [vector](<https://devfeed.tech/tags/vector.md>)

### AI overview

Token pooling reduces the number of document vectors in multi-vector retrieval, and pooling-aware training helps preserve retrieval quality after that compression. The post describes Ward clustering in Sentence Transformers: similar token vectors are merged into groups represented by their mean, while queries remain unchanged by default. Training combines the usual retrieval loss on full document embeddings with the same loss on pooled embeddings; a straight-through estimator passes gradients through pooling. In the reported NanoBEIR@13 evaluation, factor-4 pooling retained 97.8% of the trained model's unpooled nDCG@10 while using 74.8% fewer document vectors, compared with 94.6% retention for the ordinary-loss model at that budget. Adaptive thresholds can target an average vector count, though individual counts and achieved averages vary by document and corpus. The text also notes that kernel acceleration for default MaxSim scoring requires CUDA and Triton; compatible alternatives are selected where available.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Pay for Hugging Face Hub with your Google Cloud Account

DevFeed: [Pay for Hugging Face Hub with your Google Cloud Account](<https://devfeed.tech/articles/pay-for-hugging-face-hub-with-your-google-cloud-account-127705.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/hub-gcp-marketplace>)

Author: Zaïd Alhadad; Pierric Cistac; Simon Pagezy; Michelle Habonneau

Published: 2026-11-10T00:00:00Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Google Cloud Platform (GCP)](<https://devfeed.tech/topics/google-cloud.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>), [inference-endpoints](<https://devfeed.tech/topics/inference-endpoints.md>), [inference-providers](<https://devfeed.tech/topics/inference-providers.md>), [jobs](<https://devfeed.tech/topics/jobs.md>), [spaces](<https://devfeed.tech/topics/spaces.md>)

Tags: [audit-logs](<https://devfeed.tech/tags/audit-logs.md>), [billing](<https://devfeed.tech/tags/billing.md>), [compute](<https://devfeed.tech/tags/compute.md>), [gcp](<https://devfeed.tech/tags/gcp.md>), [google-cloud](<https://devfeed.tech/tags/google-cloud.md>), [google-cloud-marketplace](<https://devfeed.tech/tags/google-cloud-marketplace.md>), [guide](<https://devfeed.tech/tags/guide.md>), [hub](<https://devfeed.tech/tags/hub.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [inference-endpoints](<https://devfeed.tech/tags/inference-endpoints.md>), [inference-providers](<https://devfeed.tech/tags/inference-providers.md>), [jobs](<https://devfeed.tech/tags/jobs.md>), [linking](<https://devfeed.tech/tags/linking.md>), [marketplace](<https://devfeed.tech/tags/marketplace.md>), [models](<https://devfeed.tech/tags/models.md>), [partnerships](<https://devfeed.tech/tags/partnerships.md>)

### AI overview

Hugging Face organizations can route eligible Hub charges through a linked Google Cloud billing account by subscribing to the Hugging Face listing in Google Cloud Marketplace, choosing a project and billing account, then signing in to Hugging Face and selecting an organization where the user is an admin. The subscription requires Hugging Face review; while pending, existing billing continues. After approval, organizations with a credit card switch at the next billing period, while free-tier organizations without one usually switch immediately. The marketplace account replaces the credit card, so adding credits is unavailable. Team or Enterprise included credits apply before pay-as-you-go charges. Usage can be monitored in Hugging Face billing settings and Google Cloud Billing reports; linking and unlinking are recorded in organization audit logs.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Decision Models in Hugging Face

DevFeed: [Decision Models in Hugging Face](<https://devfeed.tech/articles/decision-models-in-hugging-face-127704.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/decision>)

Author: Aritra Roy Gosthipaty; Lysandre; Pedro Cuenca; Tom Aarsen

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [decision-making](<https://devfeed.tech/topics/decision-making.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>), [text-classification](<https://devfeed.tech/topics/text-classification.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [bernoulli](<https://devfeed.tech/tags/bernoulli.md>), [billing](<https://devfeed.tech/tags/billing.md>), [choice](<https://devfeed.tech/tags/choice.md>), [class](<https://devfeed.tech/tags/class.md>), [classes](<https://devfeed.tech/tags/classes.md>), [classification](<https://devfeed.tech/tags/classification.md>), [classifier](<https://devfeed.tech/tags/classifier.md>), [cost](<https://devfeed.tech/tags/cost.md>), [decision](<https://devfeed.tech/tags/decision.md>), [decision-models](<https://devfeed.tech/tags/decision-models.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [jev](<https://devfeed.tech/tags/jev.md>), [models](<https://devfeed.tech/tags/models.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [science](<https://devfeed.tech/tags/science.md>), [system-one](<https://devfeed.tech/tags/system-one.md>)

### AI overview

Decision models expose structured answers and probabilities directly, avoiding generated response text for tasks with predefined answers. A request pairs a state with typed questions: Choice selects among options, Score returns a probability distribution over ordered levels and its expected index, and Noul estimates the probability of a yes/no statement. The article contrasts this shared interface with fixed-label classifiers, whose classes are built into the head, and zero-shot classifiers, which accept runtime labels but focus on category assignment. It describes support in Transformers through DecisionPipeline, AutoModelForDecision, and DecisionConfig, plus a shared `/v1/systemone` serving API in Transformers and llama.cpp. Compatible autoregressive models can use batched label-logit evaluation without generating text; the article also outlines fine-tuning with Trainer and notes TRL support was still an open PR at publication time.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

DevFeed: [Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context](<https://devfeed.tech/articles/nemotron-personas-finance-japan-a-synthetic-dataset-capturing-japan-s-demographics-and-financial-context-123808.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/nvidia/nemotron-personas-finance-japan>)

Author: Atsunori Fujita; Masaya Ogushi; Vincent Gong; Kotaro Yamamoto; Yoshi Suhara; Will Jennings; Yev Meyer

Published: 2026-11-09T15:05:57Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Finance](<https://devfeed.tech/topics/finance.md>), [AI-Ready Data](<https://devfeed.tech/topics/ai-ready-data.md>), [Dataset development](<https://devfeed.tech/topics/dataset-development.md>), [Advanced AI workflows](<https://devfeed.tech/topics/advanced-ai-workflows.md>)

Tags: [advanced](<https://devfeed.tech/tags/advanced.md>), [ai](<https://devfeed.tech/tags/ai.md>), [approach](<https://devfeed.tech/tags/approach.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [context](<https://devfeed.tech/tags/context.md>), [data](<https://devfeed.tech/tags/data.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [distributions](<https://devfeed.tech/tags/distributions.md>), [domain](<https://devfeed.tech/tags/domain.md>), [finance](<https://devfeed.tech/tags/finance.md>), [financial](<https://devfeed.tech/tags/financial.md>), [financial-sector](<https://devfeed.tech/tags/financial-sector.md>), [grounded](<https://devfeed.tech/tags/grounded.md>), [japan](<https://devfeed.tech/tags/japan.md>), [japanese](<https://devfeed.tech/tags/japanese.md>), [need](<https://devfeed.tech/tags/need.md>), [nemotron](<https://devfeed.tech/tags/nemotron.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [real-world](<https://devfeed.tech/tags/real-world.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>)

### AI overview

Nemotron-Personas-Finance-Japan is a CC BY 4.0 synthetic dataset intended to support development and evaluation of financial AI for Japanese users without real customer data. It contains 9 million personas across 50 fields, pairing demographics and household context with assets, financial behavior, and literacy. The dataset was built using official statistics and public survey data, NeMo SDG-PGMs, NeMo Data Designer, and gemma-4-31B-it for Japanese text generation; industry feedback also informed its design. It adds finer education and occupation categories, underrepresented life stages, household finances, and digital and financial literacy. Proposed uses include dialogue generation, customer simulations, model evaluation, and bias testing across population groups. Personas are fully synthetic, but names follow real Japanese naming patterns; the text describes privacy risk reduction, not a guarantee of zero risk.

### Source excerpt

# Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context *A compound AI approach to Japanese personas grounded in real-world distributions* In the financial sector, the utilization of real-world data faces significant constraints due to strict privacy regulations, compliance requirements, and the need for advanced domain expertise.

## Your Experts Have Partners. There's HOPE for Keeping Them Together

DevFeed: [Your Experts Have Partners. There's HOPE for Keeping Them Together](<https://devfeed.tech/articles/your-experts-have-partners-there-s-hope-for-keeping-them-together-123807.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/amazon/higher-order-pruning-experts>)

Author: Alex M Tseng

Published: 2026-11-09T12:34:29Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>)

Tags: [activation](<https://devfeed.tech/tags/activation.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [experts](<https://devfeed.tech/tags/experts.md>), [hope](<https://devfeed.tech/tags/hope.md>), [inference](<https://devfeed.tech/tags/inference.md>), [layers](<https://devfeed.tech/tags/layers.md>), [llms](<https://devfeed.tech/tags/llms.md>), [memory](<https://devfeed.tech/tags/memory.md>), [mixture-of-experts](<https://devfeed.tech/tags/mixture-of-experts.md>), [mixture-of-experts-moe](<https://devfeed.tech/tags/mixture-of-experts-moe.md>), [mlp](<https://devfeed.tech/tags/mlp.md>), [models](<https://devfeed.tech/tags/models.md>), [moe](<https://devfeed.tech/tags/moe.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [qwen3-5](<https://devfeed.tech/tags/qwen3-5.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

HOPE reduces Mixture-of-Experts model memory by pruning experts while accounting for pairwise cooperation between them. Unlike first-order methods that score experts independently, it builds a calibration-based interaction matrix and chooses a prune set by solving a quadratic program. The post reports the best average rank among five methods across tests on three models, and a larger advantage at 50% pruning: average rank 1.58 versus 2.42 for REAP. On Qwen3.5-35B-A3B at that rate, HOPE scored 42.3 on SWE-Bench Pro, near the unpruned model's 41.8; REAP scored 36.2. Use involves one calibration pass, solving the program (reported as seconds per layer), then producing a standard smaller Hugging Face checkpoint. HOPE's edge shrinks at 10-20% pruning; it treats layers independently, and the cross-layer extension remains preliminary. The method captures pairwise, not higher-order, interactions.

### Source excerpt

Mixture-of-Experts (MoE) models have enabled state-of-the-art LLMs to be scaled to hundreds of billions of parameters, but they come with a fundamental and costly caveat: every expert MLP must occupy memory at inference time, even though the vast majority are idle for any given token. For example, Qwen3.5-122B has 256 experts in each of its 48 MoE layers, and 95% of its \~250 GB parameter memory footprint is taken up entirely by these expert MLPs alone.

## ToyBlocks: A controlled world for robot learning and evaluation

DevFeed: [ToyBlocks: A controlled world for robot learning and evaluation](<https://devfeed.tech/articles/toyblocks-a-controlled-world-for-robot-learning-and-evaluation-121632.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/allenai/toyblocks>)

Author: Kyle Wiggers

Published: 2026-11-06T14:00:48Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Robot Manipulation](<https://devfeed.tech/topics/robot-manipulation.md>), [Dataset development](<https://devfeed.tech/topics/dataset-development.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [Predictive AI](<https://devfeed.tech/topics/predictive-ai.md>), [image recognition](<https://devfeed.tech/topics/image-recognition.md>), [Isaac](<https://devfeed.tech/topics/isaac.md>), [robot sense of touch](<https://devfeed.tech/topics/robot-sense-of-touch.md>)

Tags: [code](<https://devfeed.tech/tags/code.md>), [data](<https://devfeed.tech/tags/data.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [learning](<https://devfeed.tech/tags/learning.md>), [models](<https://devfeed.tech/tags/models.md>), [robot](<https://devfeed.tech/tags/robot.md>), [world](<https://devfeed.tech/tags/world.md>)

### AI overview

ToyBlocks is an open simulation framework for studying robot learning in a controlled tabletop world, combining automatic task and demonstration generation with training and evaluation tools. A robot arm learns to push, flip, pick, place, and stack blocks; researchers can vary training data and scene properties, then test generalization on held-out tasks. Simulation ground truth helps isolate planning, execution, and prediction failures. Initial results show popular action models generalize poorly across spatial changes; ManiFlow's atomic-skill success was similar in simulation and reality (80.0% and 81.1%), while the Molmo 2-ManiFlow system completed 55.5% of simulated and 50.0% of real block-spelling episodes, with further trials needed for statistical confidence. The release includes code, models, evaluation assets, and 85,336 long-horizon demonstrations. The excerpt also reports distinct world-model results: Wan 2.2 predicted block motion and outcomes most accurately, Astra tracked the gripper but struggled with contact, and LeWM was faster and more accurate on short interactions but drifted over full trajectories.

### Source excerpt

💻 Code | 📊 Data | 🤖 Models | 📖 Documentation Robots are learning to fold laundry, sort parts, and handle unfamiliar objects. But those demonstrations are often rather narrow and don't explain what a robot has learned or why it fails. To study a complex system, scientists often begin with a small, controlled environment, such as a petri dish or terrarium, where they can vary conditions and observe the effects.

## Egyptian and Nigerian English join the Open ASR Leaderboard

DevFeed: [Egyptian and Nigerian English join the Open ASR Leaderboard](<https://devfeed.tech/articles/egyptian-and-nigerian-english-join-the-open-asr-leaderboard-121634.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/open-asr-leaderboard-liva>)

Author: Eric Bezzam; Ashley Mo; Aoi Otani

Published: 2026-11-06T00:00:00Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [announce](<https://devfeed.tech/tags/announce.md>), [asr](<https://devfeed.tech/tags/asr.md>), [audio](<https://devfeed.tech/tags/audio.md>), [australia](<https://devfeed.tech/tags/australia.md>), [canada](<https://devfeed.tech/tags/canada.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [egypt](<https://devfeed.tech/tags/egypt.md>), [english](<https://devfeed.tech/tags/english.md>), [evals](<https://devfeed.tech/tags/evals.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [highlighting](<https://devfeed.tech/tags/highlighting.md>), [india](<https://devfeed.tech/tags/india.md>), [indian](<https://devfeed.tech/tags/indian.md>), [interaction](<https://devfeed.tech/tags/interaction.md>), [languages](<https://devfeed.tech/tags/languages.md>), [leaderboard](<https://devfeed.tech/tags/leaderboard.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [open-asr-leaderboard](<https://devfeed.tech/tags/open-asr-leaderboard.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [science](<https://devfeed.tech/tags/science.md>), [source](<https://devfeed.tech/tags/source.md>), [speech](<https://devfeed.tech/tags/speech.md>)

### AI overview

The Open ASR Leaderboard has added private Egyptian- and Nigerian-accented English splits through a collaboration with Liva AI, expanding accent representation and aiming to reduce benchmark gaming with held-out data and additional providers. It will report US versus non-US results and Inner Circle (US, UK, Australia, Canada) versus Outer and Expanding Circles (India, Nigeria, Egypt); the latter two circles will be split further as more accents are added. Liva says its consumer products collect multimodal interactions from real users, yielding natural, high-context data rather than artificially elicited or task-based interactions.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

DevFeed: [New NVIDIA Cosmos 3 Open Models for Interactive World Simulation](<https://devfeed.tech/articles/new-nvidia-cosmos-3-open-models-for-interactive-world-simulation-119751.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation>)

Author: Pranjali Joshi; Rick Z; Yao Xu; Qianli Ma; Chen-Hsuan Lin; Ming-Yu Liu; Xiang Li; Zhengfei Kuang; Zhonghao Wang; Jiaxiang Tang

Published: 2026-11-05T19:35:16Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Cosmos](<https://devfeed.tech/topics/cosmos.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>), [Simulation and Design](<https://devfeed.tech/topics/simulation-and-design.md>), [Isaac](<https://devfeed.tech/topics/isaac.md>)

Tags: [cosmos](<https://devfeed.tech/tags/cosmos.md>), [cosmos3](<https://devfeed.tech/tags/cosmos3.md>), [driving](<https://devfeed.tech/tags/driving.md>), [generate](<https://devfeed.tech/tags/generate.md>), [interactive](<https://devfeed.tech/tags/interactive.md>), [lidar](<https://devfeed.tech/tags/lidar.md>), [lighting](<https://devfeed.tech/tags/lighting.md>), [model-development](<https://devfeed.tech/tags/model-development.md>), [models](<https://devfeed.tech/tags/models.md>), [nano](<https://devfeed.tech/tags/nano.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [open](<https://devfeed.tech/tags/open.md>), [platform](<https://devfeed.tech/tags/platform.md>), [sensor](<https://devfeed.tech/tags/sensor.md>), [sim](<https://devfeed.tech/tags/sim.md>), [simulation](<https://devfeed.tech/tags/simulation.md>), [stream](<https://devfeed.tech/tags/stream.md>), [view](<https://devfeed.tech/tags/view.md>), [weather](<https://devfeed.tech/tags/weather.md>), [world](<https://devfeed.tech/tags/world.md>)

### AI overview

NVIDIA's four open-weight Cosmos 3 Nano checkpoints support different simulation workflows. Transfer-Auto creates offline synchronized driving scenes across up to 11 camera views, including fisheyes, with optional joint RGB-LiDAR generation; WSM controls specify scene structure and traffic motion, while prompts guide appearance. Sim-Auto is its causally distilled streaming variant, generating successive camera and LiDAR chunks from history and updated controls. Closed-loop use also needs a simulation runtime and WSM renderer. Sim-Transfer converts simulator spatial controls such as depth, segmentation, and edges plus a text prompt into RGB observations; the simulator retains responsibility for physics and state. Sim-Bimanual predicts observations from robot and camera actions for YAM and AgiBot two-arm setups, with specific action-schema and timing conventions. The models have 16B trainable parameters and use OpenMDW1.1 terms. Inference requirements include Linux, an NVIDIA GPU, Docker, and NVIDIA Container Toolkit; the published vLLM-Omni setup was tested on one GB200, not specified as a minimum. Training guidance emphasizes validating data and held-out episodes: a short run doesa

### Source excerpt

**NVIDIA Cosmos** world model development platform adds four new open-weight checkpoints to its Cosmos 3 model family for generating interactive world simulation and multi-sensor multi-view driving scenes: - 🌦 **Cosmos3-Nano-Transfer-Auto** : Generate multi-view driving data with varied weather, lighting, and time of day. - 🚗 **Cosmos3-Nano-Sim-Auto** : Stream synchronized camera and LiDAR observations for driving simulation.

## A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

DevFeed: [A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.](<https://devfeed.tech/articles/a-closer-look-at-phonon-tts-launching-high-quality-on-device-speech-that-tops-the-benchmarks-119752.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/phonon-tts>)

Author: Olesia Krizhevich; Laurent Mazare; Eric Bezzam; Eustache Le Bihan

Published: 2026-11-05T00:00:00Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [On-device AI](<https://devfeed.tech/topics/on-device-ai.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [apache](<https://devfeed.tech/tags/apache.md>), [apache-2-0](<https://devfeed.tech/tags/apache-2-0.md>), [apple](<https://devfeed.tech/tags/apple.md>), [apple-watch](<https://devfeed.tech/tags/apple-watch.md>), [apps](<https://devfeed.tech/tags/apps.md>), [arabic](<https://devfeed.tech/tags/arabic.md>), [arm](<https://devfeed.tech/tags/arm.md>), [audio](<https://devfeed.tech/tags/audio.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [backends](<https://devfeed.tech/tags/backends.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [browser](<https://devfeed.tech/tags/browser.md>), [device](<https://devfeed.tech/tags/device.md>), [leaderboard](<https://devfeed.tech/tags/leaderboard.md>), [on-device](<https://devfeed.tech/tags/on-device.md>), [open](<https://devfeed.tech/tags/open.md>), [phonon](<https://devfeed.tech/tags/phonon.md>), [quality](<https://devfeed.tech/tags/quality.md>), [speech](<https://devfeed.tech/tags/speech.md>), [tts](<https://devfeed.tech/tags/tts.md>)

### AI overview

Phonon is a small text-to-speech model designed to generate speech locally without a server or network connection. Its 31m English model uses Apache 2.0; the 40m model supports English, French, German, Spanish, and Portuguese under a non-commercial license. The models are about 65-75 MB, include over 100 built-in voices, and support custom voices through cloning or voice design. The article reports first place in German and Spanish intelligibility and top-three placement in every evaluated language, alongside models up to 100 times larger; Phonon also leads the CPU streaming benchmark for time-to-first audio. Gradium's ptts framework targets single-stream inference across more than 10 CPU, GPU, and NPU backends. The Transformers PyTorch implementation instead targets GPU throughput through batched generation and torch.compile. The article also describes demos and integrations, and notes that leaderboard measures include intelligibility, speed, and naturalness.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Dharma Field Extractor: Any Document In, Structured Fields in JSON Out

DevFeed: [Dharma Field Extractor: Any Document In, Structured Fields in JSON Out](<https://devfeed.tech/articles/dharma-field-extractor-any-document-in-structured-fields-in-json-out-118022.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/Dharma-AI/dharma-field-extractor>)

Author: Gabriel Pimenta de Freitas Cardoso; Paulo Henrique de Medeiros Araujo; Bruno Duarte; Francisco de Almeida Rocha Alves

Published: 2026-11-04T17:50:10Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [JSON](<https://devfeed.tech/topics/json.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Amazon Textract](<https://devfeed.tech/topics/amazon-textract.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [benchmark overfitting machine learning](<https://devfeed.tech/topics/benchmark-overfitting-machine-learning.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2019](<https://devfeed.tech/tags/2019.md>), [2020](<https://devfeed.tech/tags/2020.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cost](<https://devfeed.tech/tags/cost.md>), [court](<https://devfeed.tech/tags/court.md>), [field](<https://devfeed.tech/tags/field.md>), [fields](<https://devfeed.tech/tags/fields.md>), [forms](<https://devfeed.tech/tags/forms.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [ids](<https://devfeed.tech/tags/ids.md>), [json](<https://devfeed.tech/tags/json.md>), [llm](<https://devfeed.tech/tags/llm.md>), [model](<https://devfeed.tech/tags/model.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [range](<https://devfeed.tech/tags/range.md>), [structured](<https://devfeed.tech/tags/structured.md>), [type](<https://devfeed.tech/tags/type.md>), [wide](<https://devfeed.tech/tags/wide.md>)

### AI overview

Dharma Field Extractor converts document images into typed JSON, either by finding fields without a schema or filling an exact user-provided schema; missing requested values are returned as null. The article says one model handles multiple document types without per-type fine-tuning. In the authors' 546-document, three-scenario benchmark, it scored 0.834, above four tested frontier configurations (0.741-0.816). They report costs 6-147 times lower, but compare API list-price estimates with self-hosted GPU time on one L40S, not a sales price; deployment cost depends on hardware, volume, and latency. SROIE and FUNSD were held out from training. These are the article's reported results and do not establish performance beyond its benchmark.

### Source excerpt

*A 2-billion-parameter model that extracts exactly the fields you ask for, from any document type, outperforming every frontier LLM we tested, at a fraction of the cost.* In April, we released DharmaOCR-LITE, a model specialized for OCR across a wide range of documents, from contracts and court filings to invoices, IDs and forms.

## StateBench: Testing how language models track state changes over long documents

DevFeed: [StateBench: Testing how language models track state changes over long documents](<https://devfeed.tech/articles/statebench-testing-how-language-models-track-state-changes-over-long-documents-114918.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/allenai/statebench>)

Author: Kyle Wiggers

Published: 2026-11-03T17:15:21Z

Content type: news

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [LLM Techniques](<https://devfeed.tech/topics/llm-techniques.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>)

Tags: [architecture](<https://devfeed.tech/tags/architecture.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [code](<https://devfeed.tech/tags/code.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [state](<https://devfeed.tech/tags/state.md>), [testing](<https://devfeed.tech/tags/testing.md>), [variables](<https://devfeed.tech/tags/variables.md>)

### AI overview

StateBench compares how models track changing entity states across four levels of assignment complexity, using matched sequences that vary the update type. In tests, pretrained models generally handled fixed values and direct reassignment better than cumulative or cross-variable updates; frontier models fell to 13% or lower on the longest general finite problems. When trained directly on sequences of up to 256 assignments and tested on up to 64,000, Gated DeltaNet and hybrid models generalized nearly perfectly on the three easier levels. Transformers degraded on longer sequences and did not learn general finite tasks; a more flexible Gated DeltaNet variant stayed near perfect on the hardest level to about ten times its training length. Results also varied by problem format. The benchmark assumes a fixed number of variables, and the authors say its framework does not yet cover newly appearing entities. The excerpt also notes open questions about the capabilities of constrained Gated DeltaNet variants and the effects of numerical precision and attention layers.

### Source excerpt

📄 Tech Report | 🤗 Data | 💻 Code We're introducing **StateBench**, a benchmark for *long-context state tracking*--how well language models keep track of information about entities in long documents as they evolve. State tracking underlies tasks from reading code, where variables are assigned and reassigned, to following a story as characters physically move around and objects change hands.

## Palette: teaching a small model to build decks in your design language

DevFeed: [Palette: teaching a small model to build decks in your design language](<https://devfeed.tech/articles/palette-teaching-a-small-model-to-build-decks-in-your-design-language-114920.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ibm-research/palette-deck-building>)

Author: Praveen Venkateswaran; Anupama Murthi; Hamid Adebayo

Published: 2026-11-03T16:35:36Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [lora](<https://devfeed.tech/topics/lora.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Architecture](<https://devfeed.tech/topics/ai-architecture.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Self Hosted AI Tools](<https://devfeed.tech/topics/self-hosted-ai-tools.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [rlvr](<https://devfeed.tech/topics/rlvr.md>), [grpo](<https://devfeed.tech/topics/grpo.md>)

Tags: [adapter](<https://devfeed.tech/tags/adapter.md>), [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [arithmetic](<https://devfeed.tech/tags/arithmetic.md>), [background](<https://devfeed.tech/tags/background.md>), [build](<https://devfeed.tech/tags/build.md>), [content](<https://devfeed.tech/tags/content.md>), [design](<https://devfeed.tech/tags/design.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [frontier-model](<https://devfeed.tech/tags/frontier-model.md>), [frontier-models](<https://devfeed.tech/tags/frontier-models.md>), [language](<https://devfeed.tech/tags/language.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [palette](<https://devfeed.tech/tags/palette.md>), [prompting](<https://devfeed.tech/tags/prompting.md>), [small](<https://devfeed.tech/tags/small.md>), [style](<https://devfeed.tech/tags/style.md>), [systems](<https://devfeed.tech/tags/systems.md>), [template](<https://devfeed.tech/tags/template.md>)

### AI overview

Palette is a 20B open-weight model for generating slide decks in the IBM Design Language, with data and inference designed for self-hosting. A single LoRA adapter serves two roles: a designer creates an editable JSON brief with slide content, layouts and rationale; a coder turns each slide spec into pptxgenjs code, with slide coding run in parallel. The team synthesized about 1,000 decks (roughly 10,000 slides), added counterfactual reasoning, and repaired examples using the same checks applied at inference. A harness fixes known errors, detects overlap and overflow after rendering, and keeps an editor model's rewrite only when checks show improvement. Across 131 PresentBench decks, it added 3.5 points on average. Palette ranked first in Economics and Education, third in Talk, and second across the three domains, 1.1 points behind NotebookLM. The evaluation used plans drafted by a separate frontier model; PresentBench does not measure adherence to a specific design language. The authors identify narrative decks and figure-heavy decks as areas for further work.

### Source excerpt

Ask a frontier model for a slide deck and you'll get a good one. Ask for ten and they start to look alike: the same tasteful defaults, the same house style, a faint "AI-generated deck" look. That's fine until you want the deck to look like *your* template. Frontier models have their own house style, and no amount of prompting stops them drifting back to it. And sometimes the content is confidential and can't leave your systems at all.

## Transformers Model Integration Contracts

DevFeed: [Transformers Model Integration Contracts](<https://devfeed.tech/articles/transformers-model-integration-contracts-114921.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/transformers-mic>)

Author: Tarek Ziadé

Published: 2026-11-03T00:00:00Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Transformers](<https://devfeed.tech/topics/transformers.md>), [Integration testing](<https://devfeed.tech/topics/integration-testing.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [Continuous integration](<https://devfeed.tech/topics/continuous-integration.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2026](<https://devfeed.tech/tags/2026.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [account](<https://devfeed.tech/tags/account.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [all](<https://devfeed.tech/tags/all.md>), [also](<https://devfeed.tech/tags/also.md>), [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [approach](<https://devfeed.tech/tags/approach.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [attention](<https://devfeed.tech/tags/attention.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [guide](<https://devfeed.tech/tags/guide.md>), [integration](<https://devfeed.tech/tags/integration.md>), [model](<https://devfeed.tech/tags/model.md>), [source](<https://devfeed.tech/tags/source.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

Transformers' proposed Model Integration Contracts (MICs) use small fixtures derived from real checkpoint configurations to run API usage examples on CPU during PRs. Each contract pins a fixture, a runnable Markdown usage path, and human-reviewed output baselines; CI selects contracts based on changed files and records frozen revisions and package versions. In a five-month replay, a UMT5 fixture caught an SDPA decoder masking regression from its July 30, 2026 introduction through its September 28 fix; eager attention alone missed it. A selected T5 runner step took 13.4 seconds on CPU, excluding setup. Fixtures preserve preprocessing files and original weight formats, and shrink models while retaining layer types. Random weights are preferred when responsive; overfitting a small example can improve sensitivity, though it adds maintenance. In a Whisper study, overfit fixtures caught 24 of 24 synthetic faults versus 18 of 24 for random weights, but both missed an assisted-decoding bug exposed by real weights. As of November 1, 2026, CI reports MIC failures without blocking PRs. Real-world catch rates, false alarms, review effort, real-weight-only bugs, and coverage remain unknown or,

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

[Next page](<https://devfeed.tech/sources/hugging-face-blog.md?cursor=WyIyMDI2LTExLTAzVDAwOjAwOjAwKzAwOjAwIiwgIjdhZjI1NmY0LWNmMDUtNDc3MC05MzhmLTI1OWI5NWE1MTM2OCJd>)