# Together.ai

The latest from Together

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

## Tencent Hy4 is live on Together AI

DevFeed: [Tencent Hy4 is live on Together AI](<https://devfeed.tech/articles/tencent-hy4-is-live-on-together-ai-146600.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/tencent-hy4>)

Author: Sonny Khan

Published: 2026-11-19T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [browser-game](<https://devfeed.tech/topics/browser-game.md>), [Image](<https://devfeed.tech/topics/image.md>), [prompt](<https://devfeed.tech/topics/prompt.md>)

Tags: [browser-game](<https://devfeed.tech/tags/browser-game.md>), [coding](<https://devfeed.tech/tags/coding.md>), [game](<https://devfeed.tech/tags/game.md>), [hy4](<https://devfeed.tech/tags/hy4.md>), [image](<https://devfeed.tech/tags/image.md>), [live](<https://devfeed.tech/tags/live.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [tencent](<https://devfeed.tech/tags/tencent.md>), [together-ai](<https://devfeed.tech/tags/together-ai.md>)

### AI overview

Tencent Hy4 is available on Together Serverless Inference as `tencent/Hy4`, with a 1M-token context window, image input, tool calling, and JSON output. In the described tests, it extracted all three chart values into JSON and rebuilt the chart in a single HTML file, preserving its layout and styling details except for the requested generic font. From one text prompt, it produced a playable browser game in one HTML file without libraries or external assets; a round ended and restarted successfully. Hy4 is a 770B-parameter Mixture-of-Experts model with 49B active per token. Pricing is $0.83 per 1M input tokens, $0.042 per 1M cached input tokens, and $2.50 per 1M output tokens.

### Source excerpt

Tencent Hy4 is live on Together AI. We tested its image input by extracting and rebuilding a chart, then tried its coding with a browser game generated from one prompt.

## How we make Together AI easy for coding agents to use

DevFeed: [How we make Together AI easy for coding agents to use](<https://devfeed.tech/articles/how-we-make-together-ai-easy-for-coding-agents-to-use-136086.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/how-we-do-agent-evals-at-together-ai>)

Author: Hassan El Mghari, Sam Barrowclough, Zain Hasan, Mo King

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

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [Documentation](<https://devfeed.tech/topics/documentation.md>), [Test automation](<https://devfeed.tech/topics/test-automation.md>), [Agent Skills](<https://devfeed.tech/topics/agent-skills.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>)

Tags: [accessibility](<https://devfeed.tech/tags/accessibility.md>), [accessible](<https://devfeed.tech/tags/accessible.md>), [agent-evals](<https://devfeed.tech/tags/agent-evals.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [api](<https://devfeed.tech/tags/api.md>), [coding](<https://devfeed.tech/tags/coding.md>), [coding-agents](<https://devfeed.tech/tags/coding-agents.md>), [docs](<https://devfeed.tech/tags/docs.md>), [errors](<https://devfeed.tech/tags/errors.md>), [evals](<https://devfeed.tech/tags/evals.md>), [skills](<https://devfeed.tech/tags/skills.md>), [together-ai](<https://devfeed.tech/tags/together-ai.md>)

### AI overview

Together AI built coding-agent evaluations from realistic developer prompts, running each task across baseline, documentation MCP, and skill configurations with multiple models and attempts. Session logs and an LLM judge helped identify where agents failed or recovered. A GPU pricing task exposed misleading CLI sample output and a skill link that omitted pricing guidance; adding two sentences directing agents to the live regions API and pricing table raised that skill-enabled task's pass rate from 4 of 15 Sonnet attempts to 100% across tested models. The task also uncovered an API error that pointed to a nonexistent route and a retry routing bug, both of which were fixed. After consolidating 14 skills, the initial unified version passed 75% of runs and cost over twice as much; shortening its router, fixing stale information and sample-code bugs, and adding maintenance automation raised results to 70 of 72 runs, with 20% less runtime than the original skills and costs back to baseline. Overall, guided configurations reached 96% for docs MCP and 97% for the unified skill, compared with an 82% baseline in the September snapshot. These are not controlled comparisons: snapshots differed

### Source excerpt

We built agent evals to test how effective coding agents are at using Together AI. Small fixes to our docs, skills, and API errors lifted task completion rate from 75% to 97%.

## Making open models work for long-horizon agentic workloads

DevFeed: [Making open models work for long-horizon agentic workloads](<https://devfeed.tech/articles/making-open-models-work-for-long-horizon-agentic-workloads-138313.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/making-open-models-work-for-long-horizon-agentic-workloads>)

Author: Zain Hasan, Parth Sareen (Ollama)

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

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [Ollama](<https://devfeed.tech/topics/ollama.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-apps](<https://devfeed.tech/tags/agentic-apps.md>), [agentic-workloads](<https://devfeed.tech/tags/agentic-workloads.md>), [agents](<https://devfeed.tech/tags/agents.md>), [api](<https://devfeed.tech/tags/api.md>), [cache](<https://devfeed.tech/tags/cache.md>), [caching](<https://devfeed.tech/tags/caching.md>), [calendar](<https://devfeed.tech/tags/calendar.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [coding](<https://devfeed.tech/tags/coding.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [harness](<https://devfeed.tech/tags/harness.md>), [models](<https://devfeed.tech/tags/models.md>), [ollama](<https://devfeed.tech/tags/ollama.md>), [open](<https://devfeed.tech/tags/open.md>), [session](<https://devfeed.tech/tags/session.md>), [together-ai](<https://devfeed.tech/tags/together-ai.md>)

### AI overview

Agent workloads resend a growing conversation with each turn, so prefix caching can reduce repeated prompt processing, but only when both the harness preserves the prompt prefix and the inference engine retains and routes to the cached session. The harness should keep system prompts and tool definitions stable, place changing session details in the task message, choose reasoning effort early, disclose when tool output is truncated, and avoid giving subagents irrelevant history. Late-loaded tool definitions can preserve the prefix when appended, though some chat templates do not support them there. On the engine side, routing may use a session key or the longest cached prefix; load can make a cache miss preferable to waiting on an overloaded node. Engines can offload idle caches to CPU memory, use paged allocation and continuous batching, and separate prefill from decode. Caching avoids recomputing old prompt work, but cached tokens still consume memory and are attended to during generation.

### Source excerpt

Every agent turn resends the full session. Together AI and Ollama break down how to protect the prefix cache, from the harness to the GPU.

## Together Compiler: Continual Compilation for AI inference

DevFeed: [Together Compiler: Continual Compilation for AI inference](<https://devfeed.tech/articles/together-compiler-continual-compilation-for-ai-inference-118999.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/together-compiler-continual-compilation-for-ai-inference>)

Author: Reyna Abhyankar, Dylan Lim, and Simran Arora

Published: 2026-11-04T00:00:00Z

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [Compiler](<https://devfeed.tech/topics/compiler.md>), [agent orchestration](<https://devfeed.tech/topics/agent-orchestration.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-systems](<https://devfeed.tech/tags/agentic-systems.md>), [agentic-workloads](<https://devfeed.tech/tags/agentic-workloads.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [ai-systems](<https://devfeed.tech/tags/ai-systems.md>), [ai-workloads](<https://devfeed.tech/tags/ai-workloads.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [attention](<https://devfeed.tech/tags/attention.md>), [compilation](<https://devfeed.tech/tags/compilation.md>), [compiler](<https://devfeed.tech/tags/compiler.md>), [engines](<https://devfeed.tech/tags/engines.md>), [inference](<https://devfeed.tech/tags/inference.md>), [models](<https://devfeed.tech/tags/models.md>), [optimize](<https://devfeed.tech/tags/optimize.md>), [throughput](<https://devfeed.tech/tags/throughput.md>), [traffic](<https://devfeed.tech/tags/traffic.md>)

### AI overview

Together Compiler (TKCC) targets changing inference workloads by using parallel agents and persistent engines that hot-reload candidate forward passes, avoiding the tens-of-minutes startup and teardown costs of conventional engines. Fast engines support rapid, low-variance experiments; slower engines test optimizations on more layers and realistic data distributions. A tiered cost model combines hardware specifications, microbenchmarks, and measured passes to prune parallelism choices while accounting for traffic properties such as MoE imbalance. The excerpt reports up to 2.6x higher throughput across frontier open models in under 24 hours, 1.7x on Llama 1B at batch size 1, and low-latency results of 1,111 TPS on MiniMax M3.1-Flash and 1,205 TPS on DeepSeek v4.1-Flash. It also reports 252x higher experiment throughput than baseline agentic optimization systems and generalization to several non-LLM workloads. A footnote says the Trainium2 result used synthetic traffic because higher concurrencies were not supported.

### Source excerpt

Together Compiler uses agents and persistent engines to optimize open models in under 24 hours, delivering up to 3.7x higher throughput as traffic changes.

## Polymorphic Inference: self-tuning inference for open models

DevFeed: [Polymorphic Inference: self-tuning inference for open models](<https://devfeed.tech/articles/polymorphic-inference-self-tuning-inference-for-open-models-118998.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/polymorphic-inference>)

Author: By Simran Arora (SVP of Research, AI Systems) and Dan Fu (Founding SVP of Research)

Published: 2026-11-04T00:00:00Z

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [model-serving](<https://devfeed.tech/topics/model-serving.md>), [vllm](<https://devfeed.tech/topics/vllm.md>)

Tags: [2025](<https://devfeed.tech/tags/2025.md>), [2026](<https://devfeed.tech/tags/2026.md>), [about](<https://devfeed.tech/tags/about.md>), [advanced](<https://devfeed.tech/tags/advanced.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-applications](<https://devfeed.tech/tags/ai-applications.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [ai-powered](<https://devfeed.tech/tags/ai-powered.md>), [chips](<https://devfeed.tech/tags/chips.md>), [compiler](<https://devfeed.tech/tags/compiler.md>), [inference](<https://devfeed.tech/tags/inference.md>), [keep](<https://devfeed.tech/tags/keep.md>), [models](<https://devfeed.tech/tags/models.md>), [open](<https://devfeed.tech/tags/open.md>), [production](<https://devfeed.tech/tags/production.md>), [traffic](<https://devfeed.tech/tags/traffic.md>), [uses](<https://devfeed.tech/tags/uses.md>)

### AI overview

Together AI argues that inference systems must adapt as open models, architectures, traffic patterns and hardware change. It says static engine configurations and kernels tuned for fixed workloads may not perform well under different production conditions; even megakernel fusion is not a universal goal for this more varied setting. Its Together Compiler reframes optimization as agent-guided search over performance patterns using a persistent engine, and currently runs continuous optimization on some popular production serverless endpoints. The company says this lets it adapt automatically to new traffic shapes and workloads, while dedicated-endpoint self-service optimization is planned and available to customers only in limited beta. Other planned research includes workload-adaptive speculative decoding and routing across heterogeneous deployments. These are the company's claims and roadmap; the excerpt provides no quantitative performance results.

### Source excerpt

Models, traffic and chips are constantly changing. Polymorphic Inference adapts to each, and Together Compiler uses AI agents to keep production inference tuned.

## Auto repair for Together GPU Clusters: from GPU fault to healthy node

DevFeed: [Auto repair for Together GPU Clusters: from GPU fault to healthy node](<https://devfeed.tech/articles/auto-repair-for-together-gpu-clusters-from-gpu-fault-to-healthy-node-103670.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/auto-repair-for-together-gpu-clusters>)

Author: Pavneet Ahluwalia, Dongsheng Zhang, Barbara Kazanowska

Published: 2026-10-27T00:00:00Z

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Distributed Systems & Parallel Computing](<https://devfeed.tech/topics/distributed-systems-parallel-computing.md>), [NCCL](<https://devfeed.tech/topics/nccl.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [account](<https://devfeed.tech/tags/account.md>), [actions](<https://devfeed.tech/tags/actions.md>), [alert](<https://devfeed.tech/tags/alert.md>), [alerts](<https://devfeed.tech/tags/alerts.md>), [also](<https://devfeed.tech/tags/also.md>), [annotation](<https://devfeed.tech/tags/annotation.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [automatically](<https://devfeed.tech/tags/automatically.md>), [clusters](<https://devfeed.tech/tags/clusters.md>), [code](<https://devfeed.tech/tags/code.md>), [fabric](<https://devfeed.tech/tags/fabric.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [gpu-clusters](<https://devfeed.tech/tags/gpu-clusters.md>), [node](<https://devfeed.tech/tags/node.md>), [repair](<https://devfeed.tech/tags/repair.md>), [run](<https://devfeed.tech/tags/run.md>), [works](<https://devfeed.tech/tags/works.md>)

### AI overview

Together GPU Clusters' auto-repair pipeline uses failed health checks and supporting evidence to recommend the lightest mapped fix; `NotReady` alone never triggers repair. Detection runs inside VMs and on hosts, while policy controls which faults run automatically and how long repairs wait for workloads to become idle. Repairs proceed through cordon, wait, drain, action and rejoin, with escalation after repeated failures, evidence and capacity gates for migrations, and safeguards against repair storms. Customers can submit repairs through the API or console; node-label remediation contracts are in preview. After migration or removal, the vacated machine is automatically tested and returned to service, ticketed after explicit failure, or restored to its prior state if validation is aborted. Reboots preserve local scratch data; quick reprovisioning loses it. The excerpt also describes host power cycling as a preview feature.

### Source excerpt

Together GPU Clusters auto repair detects GPU, host and fabric faults, picks the lightest fix, and lets you decide per fault and per Xid code which repairs run automatically. Here is how it works.

## Together launches a service for custom post-training of open models

DevFeed: [Together launches a service for custom post-training of open models](<https://devfeed.tech/articles/together-post-training-frontier-grade-model-shaping-the-way-that-fits-you-90284.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/together-post-training>)

Author: Ivan Provilkov, Gleb Khaykin, Denis Mazur, Nikita Smetanin, Egor Timofeev, Conner Manuel, Felix Zhang, Alex Mazitov, Jasmine Li, Gleb Vazhenin, Sergei Vorobyov, Qingyang Wu, Derek Petersen, Arseniy Be

Published: 2026-10-19T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [post-training](<https://devfeed.tech/topics/post-training.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [Advanced AI workflows](<https://devfeed.tech/topics/advanced-ai-workflows.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [distributed-training](<https://devfeed.tech/topics/distributed-training.md>), [peft](<https://devfeed.tech/topics/peft.md>)

Tags: [agent-harness](<https://devfeed.tech/tags/agent-harness.md>), [glm](<https://devfeed.tech/tags/glm.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [kimi](<https://devfeed.tech/tags/kimi.md>), [llms](<https://devfeed.tech/tags/llms.md>), [model](<https://devfeed.tech/tags/model.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [rl](<https://devfeed.tech/tags/rl.md>), [sft](<https://devfeed.tech/tags/sft.md>), [train](<https://devfeed.tech/tags/train.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

Together introduces a service for post-training open models with supervised fine-tuning or reinforcement learning. Customers define the training loop and experiment setup, while Together manages the underlying infrastructure and offers dedicated GPU resources, multiple LoRA adapters, and support for training and deployment workflows.

### Source excerpt

Post-train open LLMs like Qwen, GLM and Kimi into custom models you own, with SFT or RL. You write the training loop. Together runs the hard infrastructure.

## Meta Muse Spark Open is live on Together AI

DevFeed: [Meta Muse Spark Open is live on Together AI](<https://devfeed.tech/articles/meta-muse-spark-open-is-live-on-together-ai-88079.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/meta-muse-spark-open>)

Published: 2026-10-15T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [gpt-oss](<https://devfeed.tech/topics/gpt-oss.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [agentic-engineering](<https://devfeed.tech/tags/agentic-engineering.md>), [agentic-software](<https://devfeed.tech/tags/agentic-software.md>), [ai](<https://devfeed.tech/tags/ai.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [meta](<https://devfeed.tech/tags/meta.md>), [ollama](<https://devfeed.tech/tags/ollama.md>), [serverless](<https://devfeed.tech/tags/serverless.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

Together AI announces that Meta Muse Spark Open, an open-weight model for agentic coding and other multimodal workflows, is available on its platform. The release highlights a 1M-token context window, long-running autonomous task support, computer use, browsing, and multimodal input. Early access customers Factory and Ollama describe evaluating it for tool calling and long-context work; Together AI lists serverless API pricing.

### Source excerpt

Meta Muse Spark Open, Meta's open-weight model for agentic coding, is live on Together AI with a 1M-token context window on serverless and dedicated infrastructure.

## Expanding our enterprise inference capacity with IBM Cloud and NVIDIA

DevFeed: [Expanding our enterprise inference capacity with IBM Cloud and NVIDIA](<https://devfeed.tech/articles/expanding-our-enterprise-inference-capacity-with-ibm-cloud-and-nvidia-80176.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/expanding-our-enterprise-inference-capacity-with-ibm-cloud-and-nvidia>)

Author: Kai Mak

Published: 2026-10-06T00:00:00Z

Content type: news

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [NVIDIA Vera Rubin](<https://devfeed.tech/topics/nvidia-vera-rubin.md>), [Nvidia](<https://devfeed.tech/topics/nvidia.md>), [NVIDIA Spectrum-X Ethernet](<https://devfeed.tech/topics/nvidia-spectrum-x-ethernet.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [Cloud](<https://devfeed.tech/topics/cloud.md>)

Tags: [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [developers](<https://devfeed.tech/tags/developers.md>), [ethernet](<https://devfeed.tech/tags/ethernet.md>), [gpus](<https://devfeed.tech/tags/gpus.md>), [ibm](<https://devfeed.tech/tags/ibm.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

Together AI, IBM, and NVIDIA are collaborating on a dedicated NVIDIA B300 inference cluster on IBM Cloud, with Together AI operating the inference layer. The announcement describes the cluster as a large-scale production infrastructure for serving open models, supported by Spectrum-X Ethernet networking.

### Source excerpt

Enterprises can now run open models at production scale on a dedicated B300 inference cluster, built by Together AI, IBM Cloud, and NVIDIA

## Together Link routes coding agent tasks to open models with session cost tracking

DevFeed: [Together Link routes coding agent tasks to open models with session cost tracking](<https://devfeed.tech/articles/together-link-open-models-in-the-harness-you-already-use-start-with-one-command-today-80226.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/together-link-frontier-quality-open-models-in-the-harness-you-already-use>)

Author: Will Van Eaton, Hassan El Mghari

Published: 2026-10-05T00:00:00Z

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [DeepSeek Harness](<https://devfeed.tech/topics/dsh.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [coding-agent](<https://devfeed.tech/tags/coding-agent.md>), [coding-agents](<https://devfeed.tech/tags/coding-agents.md>), [cost](<https://devfeed.tech/tags/cost.md>), [developers](<https://devfeed.tech/tags/developers.md>), [glm](<https://devfeed.tech/tags/glm.md>), [glm-5-3](<https://devfeed.tech/tags/glm-5-3.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [kimi](<https://devfeed.tech/tags/kimi.md>), [kimi-k3](<https://devfeed.tech/tags/kimi-k3.md>), [opencode](<https://devfeed.tech/tags/opencode.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [opus-5-5](<https://devfeed.tech/tags/opus-5-5.md>), [platform](<https://devfeed.tech/tags/platform.md>), [pricing](<https://devfeed.tech/tags/pricing.md>)

### AI overview

Together Link connects supported coding agents to Together AI models, including GLM 5.3 and Kimi K3. Its Auto router selects a model based on the session's first task, and a tracker compares each session's cost with what it would have cost on Opus 5.5. The article says this can reduce model spending while letting engineering teams continue using familiar tools.

### Source excerpt

Together Link brings frontier open models like GLM 5.3 and Kimi K3 into the coding agent your team already uses, cutting model spend by over 50%.

## How to train your own Jev for $17

DevFeed: [How to train your own Jev for $17](<https://devfeed.tech/articles/how-to-train-your-own-jev-for-17-80186.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/how-to-train-your-own-jev>)

Author: Hassan El Mghari

Published: 2026-09-23T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [post-training](<https://devfeed.tech/topics/post-training.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [Cloudflare Workers](<https://devfeed.tech/topics/cloudflare-workers.md>), [rlhf](<https://devfeed.tech/topics/rlhf.md>), [argilla](<https://devfeed.tech/topics/argilla.md>)

Tags: [chat-template](<https://devfeed.tech/tags/chat-template.md>), [cli](<https://devfeed.tech/tags/cli.md>), [cost](<https://devfeed.tech/tags/cost.md>), [deploy](<https://devfeed.tech/tags/deploy.md>), [environment-variables](<https://devfeed.tech/tags/environment-variables.md>), [errors](<https://devfeed.tech/tags/errors.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [qwen3-5](<https://devfeed.tech/tags/qwen3-5.md>), [serverless](<https://devfeed.tech/tags/serverless.md>)

### AI overview

A step-by-step guide to fine-tuning a Jev-like classification model with Qwen3.5 4B using Together AI. It covers preparing datasets, training and deploying the model to an endpoint, then querying it for structured classification results.

### Source excerpt

We just launched our own Jev-like classifier, together/Tev1-4B-experimental, on top of Qwen3.5 4B on Together's serverless platform. In this blog post we'll show you how to fine-tune your own version!

## Canary rollouts: upgrade models in production without downtime

DevFeed: [Canary rollouts: upgrade models in production without downtime](<https://devfeed.tech/articles/canary-rollouts-upgrade-models-in-production-without-downtime-80164.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/canary-rollouts-upgrade-models-in-production-without-downtime>)

Author: Siddhant Bharti, Zarni Phyo, Zain Hasan, Nikitha Suryadevara, Ted Cui

Published: 2026-09-22T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [blue green and canary deployments](<https://devfeed.tech/topics/blue-green-and-canary-deployments.md>), [blue-green deployment](<https://devfeed.tech/topics/blue-green-deployment.md>), [zero downtime deployment](<https://devfeed.tech/topics/zero-downtime-deployment.md>), [progressive delivery](<https://devfeed.tech/topics/progressive-delivery.md>), [inference-providers](<https://devfeed.tech/topics/inference-providers.md>)

Tags: [canary](<https://devfeed.tech/tags/canary.md>), [dedicated-inference](<https://devfeed.tech/tags/dedicated-inference.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [inference](<https://devfeed.tech/tags/inference.md>), [production](<https://devfeed.tech/tags/production.md>), [rollback](<https://devfeed.tech/tags/rollback.md>), [traffic](<https://devfeed.tech/tags/traffic.md>)

### AI overview

This guide explains how staged canary, blue-green, and rolling deployments move live inference traffic between model deployments. It covers health checks, metric gates, traffic and capacity safeguards, API and CLI controls, and a live Qwen model upgrade that paused after a latency regression and was reversed without failed requests.

### Source excerpt

A hard model swap exposes every user at once, and rolling back means cold-starting the old deployment under pressure. Here's how staged traffic ramps, metric gates, and automatic rollback work on dedicated inference.

## How a global fintech scaled coding agent traffic with Dedicated Model Inference

DevFeed: [How a global fintech scaled coding agent traffic with Dedicated Model Inference](<https://devfeed.tech/articles/how-a-global-fintech-scaled-coding-agent-traffic-with-dedicated-model-inference-80183.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/global-fintech-scales-coding-agent-traffic-with-dedicated-model-inference>)

Author: Ankit Goyal

Published: 2026-09-18T00:00:00Z

Content type: article

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [model-serving](<https://devfeed.tech/topics/model-serving.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [friendliai](<https://devfeed.tech/topics/friendliai.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>)

Tags: [ai-coding-agents](<https://devfeed.tech/tags/ai-coding-agents.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [concurrency](<https://devfeed.tech/tags/concurrency.md>), [fintech](<https://devfeed.tech/tags/fintech.md>), [inference](<https://devfeed.tech/tags/inference.md>)

### AI overview

A global fintech moved its coding assistant to Together's Dedicated Model Inference platform to handle bursty engineering-hour traffic and give its teams direct control over endpoint scaling, model updates, and testing. The article describes how observability helped diagnose request queuing and how the team tuned routing and capacity for concurrency. The customer is also scoping a second workload in a new region.

### Source excerpt

Inside a global bank's shift to self-serve dedicated inference: how Together's DMI gave engineering teams direct control over scaling, models, and testing.

## Migrating from closed to open source models, Together

DevFeed: [Migrating from closed to open source models, Together](<https://devfeed.tech/articles/migrating-from-closed-to-open-source-models-together-80199.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/migrating-from-closed-to-open-source-models>)

Author: Kareem Abdol-Hamid, Zain Hasan, Sonny Khan

Published: 2026-09-16T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

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

Tags: [canary-deployments](<https://devfeed.tech/tags/canary-deployments.md>), [context-engineering](<https://devfeed.tech/tags/context-engineering.md>), [cost](<https://devfeed.tech/tags/cost.md>), [databases](<https://devfeed.tech/tags/databases.md>), [open-source](<https://devfeed.tech/tags/open-source.md>)

### AI overview

The article outlines a five-stage approach for moving software workloads from closed models to open-source models: discover candidates, evaluate them with relevant benchmarks and real traffic, adapt prompts and model settings, assess effort, risks, and return, then roll out through a production roadmap. It emphasizes workload-specific evaluations and incremental validation, including canary deployments.

### Source excerpt

Moving from closed to open source models can take weeks, not years. A five-stage playbook: discover, evaluate, adapt, decide, and production.

## Together AI expands fine-tuning service with more models, live metrics, and finer controls

DevFeed: [Together AI expands fine-tuning service with more models, live metrics, and finer controls](<https://devfeed.tech/articles/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-80223.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls>)

Author: Artem Chumachenko, Egor Timofeev, Jasmine Li, Ruslan Khaidurov, Nikita Smetanin, Sergei Vorobyov, Arseniy Belorukov, Denis Fedorenko, Alex Moldovan, Shadi Mokhtar, Gleb Vazhenin, Sonny Khan, Adee Fein

Published: 2026-09-11T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [lora](<https://devfeed.tech/topics/lora.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [experiment-tracking](<https://devfeed.tech/tags/experiment-tracking.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>)

### AI overview

Together Fine-Tuning adds support for new open-weight models and expands controls for the model training workflow. Updates include live training metrics and run comparisons, Expert LoRA for mixture-of-experts layers, early stopping, configurable gradient accumulation and sequence packing, tokenized data previews, sample weights, and pre-flight dataset validation. Training prices for selected models are also reduced.

### Source excerpt

Together Fine-Tuning adds the latest open-weight models, live experiment tracking, Expert LoRA, early stopping, tokenized dataset previews, pre-flight validation, and lower training prices on selected models.

## To Infinity and Beyond: ThunderKittens Now on NVIDIA Vera Rubin NVL72!

DevFeed: [To Infinity and Beyond: ThunderKittens Now on NVIDIA Vera Rubin NVL72!](<https://devfeed.tech/articles/to-infinity-and-beyond-thunderkittens-now-on-nvidia-vera-rubin-nvl72-80217.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/to-infinity-and-beyond-thunderkittens-now-on-nvidia-vera-rubin-nvl72>)

Author: Dylan Lim, Xinyi Li, Sonny Li, Peter Wu, Dan Fu, Simran Arora

Published: 2026-09-10T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Vera Rubin NVL72](<https://devfeed.tech/topics/vera-rubin-nvl72.md>), [NVIDIA Vera Rubin](<https://devfeed.tech/topics/nvidia-vera-rubin.md>), [GPU](<https://devfeed.tech/topics/gpu.md>), [NVFP4](<https://devfeed.tech/topics/nvfp4.md>), [High-Performance Computing](<https://devfeed.tech/topics/high-performance-computing.md>)

Tags: [blackwell](<https://devfeed.tech/tags/blackwell.md>), [cuda](<https://devfeed.tech/tags/cuda.md>), [nvfp4](<https://devfeed.tech/tags/nvfp4.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [nvl72](<https://devfeed.tech/tags/nvl72.md>), [vera-rubin](<https://devfeed.tech/tags/vera-rubin.md>), [vera-rubin-nvl72](<https://devfeed.tech/tags/vera-rubin-nvl72.md>)

### AI overview

Together's kernels team describes porting ThunderKittens GEMM kernels to NVIDIA Vera Rubin NVL72. It explains how the platform's wider tensor-core K step, additional tensor memory and shared memory, B-side collector, and early A release affect kernel design. The team reports that its optimizations bring an NVFP4 GEMM to over 22 PFLOPS and competitive performance with cuBLAS and CuTe DSL.

### Source excerpt

We ported ThunderKittens to NVIDIA's Vera Rubin NVL72 and rebuilt our NVFP4 GEMM around the new hardware, taking it from 42% of roofline to over 22 PFLOPS -- competitive with cuBLAS and CuTe DSL. Here is what changed in the ISA and how we used it.

## Introducing preemptible compute: the same compute, half the price

DevFeed: [Introducing preemptible compute: the same compute, half the price](<https://devfeed.tech/articles/introducing-preemptible-compute-the-same-compute-half-the-price-80191.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/introducing-preemptible-compute-the-same-compute-half-the-price>)

Author: Pavneet Ahluwalia, Varun Chakravarthy

Published: 2026-09-10T00:00:00Z

Content type: release

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [autoscaling](<https://devfeed.tech/topics/autoscaling.md>), [jobs](<https://devfeed.tech/topics/jobs.md>), [Amazon EC2](<https://devfeed.tech/topics/amazon-ec2.md>)

Tags: [ai-compute](<https://devfeed.tech/tags/ai-compute.md>), [cost](<https://devfeed.tech/tags/cost.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [gpu-clusters](<https://devfeed.tech/tags/gpu-clusters.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [kubernetes](<https://devfeed.tech/tags/kubernetes.md>)

### AI overview

Together introduces preemptible compute in public preview for its Kubernetes GPU clusters. The nodes use spare capacity at a fixed 50% of the on-demand rate and can be reclaimed with up to five minutes for workloads to checkpoint and exit. The release targets interruptible jobs such as experiments, batch inference, and temporary bursts; critical components are meant to stay on standard nodes.

### Source excerpt

Together GPU Clusters now supports preemptible compute: the same GPU capacity at a flat 50% of the on-demand rate, with a five-minute drain window.

## The Open Source AI Stack

DevFeed: [The Open Source AI Stack](<https://devfeed.tech/articles/the-open-source-ai-stack-80214.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/the-open-source-ai-stack>)

Author: Hassan El Mghari

Published: 2026-09-09T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>), [litellm](<https://devfeed.tech/topics/litellm.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-coding](<https://devfeed.tech/tags/ai-coding.md>), [model-inference](<https://devfeed.tech/tags/model-inference.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-source-ai](<https://devfeed.tech/tags/open-source-ai.md>), [workflow](<https://devfeed.tech/tags/workflow.md>)

### AI overview

The article explains how developers can assemble an open model AI stack from independent layers: a model, inference provider, gateway or router, coding harness, and tools such as skills and MCP servers. It describes tradeoffs between larger and smaller models and recommends experimenting with combinations for coding workflows, while using separate sessions and planning, implementation, and review stages to manage context and work.

### Source excerpt

A deep dive into the open model AI stack -- model, inference, gateways and routers, harness, and tools -- and how keeping each layer independent lets you swap in a new open model in minutes instead of rebuilding your workflow.

## GLM-5.3 vs. GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

DevFeed: [GLM-5.3 vs. GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing](<https://devfeed.tech/articles/glm-5-3-vs-gpt-5-6-sol-on-deepswe-cost-coding-and-routing-80182.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/glm-5-3-vs-gpt-5-6-sol-on-deepswe-cost-coding-and-routing>)

Author: Zain Hasan, Shobhit Dixit

Published: 2026-08-21T00:00:00Z

Content type: comparison

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [devstral](<https://devfeed.tech/topics/devstral.md>), [gpt-6-astra](<https://devfeed.tech/topics/gpt-6-astra.md>), [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [cost](<https://devfeed.tech/tags/cost.md>), [deepswe-v1-1](<https://devfeed.tech/tags/deepswe-v1-1.md>), [glm](<https://devfeed.tech/tags/glm.md>), [glm-5-3](<https://devfeed.tech/tags/glm-5-3.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-5-6](<https://devfeed.tech/tags/gpt-5-6.md>), [gpt-5-6-sol](<https://devfeed.tech/tags/gpt-5-6-sol.md>), [routing](<https://devfeed.tech/tags/routing.md>)

### AI overview

A 904 rollout comparison of GLM-5.3 and GPT-5.6 Sol on the DeepSWE software engineering benchmark finds Sol slightly ahead on single attempts, while GLM-5.3 leads across retries at about half the per-rollout cost. The article also examines speed, failure patterns, programming language and task-domain results, and proposes a GLM-first cascade that escalates to Sol when tests reject an answer; that setup reaches 85.9% at $6.61 per task in this run.

### Source excerpt

We ran 904 DeepSWE rollouts on GLM-5.3 and GPT-5.6 Sol. Sol leads pass@1 by 3.7 points; GLM-5.3 wins pass@4 at half the cost, and a GLM-first cascade hits 85.9%.

## GLM-5.3 vs. Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

DevFeed: [GLM-5.3 vs. Claude Fable 5 on DeepSWE: Cost, Coding, and Routing](<https://devfeed.tech/articles/glm-5-3-vs-claude-fable-5-on-deepswe-cost-coding-and-routing-80181.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/glm-5-3-vs-claude-fable-5-on-deepswe-cost-coding-and-routing>)

Author: Zain Hasan; Shobhit Dixit

Published: 2026-08-21T00:00:00Z

Content type: comparison

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [DeepSeek Harness](<https://devfeed.tech/topics/dsh.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [claude](<https://devfeed.tech/tags/claude.md>), [coding](<https://devfeed.tech/tags/coding.md>), [cost](<https://devfeed.tech/tags/cost.md>), [glm-5-3](<https://devfeed.tech/tags/glm-5-3.md>), [routing](<https://devfeed.tech/tags/routing.md>), [vs](<https://devfeed.tech/tags/vs.md>)

### AI overview

The article compares GLM-5.3 and Claude Fable 5 on 113 DeepSWE software engineering tasks. Their first-attempt accuracy is close, while GLM-5.3 scores higher with retries and costs $3.99 per rollout compared with Fable 5 at $21.63. It recommends GLM-5.3 as the default and Fable 5 for Rust-heavy or serialization-critical tasks, while noting that the models often succeed and fail on the same tasks.

### Source excerpt

We ran 904 DeepSWE rollouts on GLM-5.3 and Claude Fable 5. A tie on pass@1, but GLM-5.3 wins pass@4 and costs 5.4x less: \$3.99 per rollout vs. \$21.63.

## DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

DevFeed: [DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and Routing](<https://devfeed.tech/articles/deepseek-v4-pro-0813-vs-claude-fable-5-on-deepswe-cost-coding-and-routing-80169.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/deepseek-v4-pro-0813-vs-claude-fable-5-on-deepswe-cost-coding-and-routing>)

Author: Zain Hasan, Shobhit Dixit

Published: 2026-08-17T00:00:00Z

Content type: comparison

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [DeepSeek Harness](<https://devfeed.tech/topics/dsh.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [Test-driven development](<https://devfeed.tech/topics/tdd.md>)

Tags: [claude-fable-5](<https://devfeed.tech/tags/claude-fable-5.md>), [coding](<https://devfeed.tech/tags/coding.md>), [concurrency](<https://devfeed.tech/tags/concurrency.md>), [cost](<https://devfeed.tech/tags/cost.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>)

### AI overview

The comparison reports DeepSeek V4 Pro 0813 at 62.8% pass@1 and Claude Fable 5 at 69.7%, while Pro reaches 88.5% pass@4 versus Fable's 84.1% at about 90 times lower cost per rollout. It argues that running Pro first and escalating to Fable when tests reject the result solves 82.7% of tasks at $8.28 per task in this benchmark run.

### Source excerpt

We ran 904 DeepSWE rollouts on DeepSeek V4 Pro 0813 and Claude Fable 5. Fable leads pass@1 at 90x the cost; Pro wins pass@4, and a Pro-first cascade hits 82.7%.

## A/B test models in production

DevFeed: [A/B test models in production](<https://devfeed.tech/articles/a-b-test-models-in-production-80154.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/a-b-test-models-in-production>)

Author: Zain Hasan, Zarni Phyo, Nikitha Suryadevara, Ted Cui

Published: 2026-08-17T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [A/B Testing](<https://devfeed.tech/topics/a-b-testing.md>), [model-deployment](<https://devfeed.tech/topics/model-deployment.md>), [blue green and canary deployments](<https://devfeed.tech/topics/blue-green-and-canary-deployments.md>)

Tags: [a-b-testing](<https://devfeed.tech/tags/a-b-testing.md>), [ai-platform](<https://devfeed.tech/tags/ai-platform.md>), [api](<https://devfeed.tech/tags/api.md>), [deployment](<https://devfeed.tech/tags/deployment.md>)

### AI overview

The article explains how to use Together AI endpoint-level A/B experiments to compare candidate language models with real user traffic. It covers fixed traffic shares, cohort attribution, ramping variants, measuring product outcomes alongside platform metrics, and promoting or removing a variant.

### Source excerpt

Shadow traffic proves a candidate is operationally sound. It can't tell you if users like it better. Run the split at the endpoint instead of in your app code.

## DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding

DevFeed: [DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding](<https://devfeed.tech/articles/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding-107455.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding>)

Author: Zain Hasan, Shobhit Dixit

Published: 2026-08-06T00:00:00Z

Content type: comparison

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [Test automation](<https://devfeed.tech/topics/test-automation.md>)

Tags: [coding](<https://devfeed.tech/tags/coding.md>), [cost](<https://devfeed.tech/tags/cost.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepswe](<https://devfeed.tech/tags/deepswe.md>), [flash](<https://devfeed.tech/tags/flash.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-5](<https://devfeed.tech/tags/gpt-5.md>), [gpt-5-6](<https://devfeed.tech/tags/gpt-5-6.md>), [luna](<https://devfeed.tech/tags/luna.md>), [pass](<https://devfeed.tech/tags/pass.md>), [vs](<https://devfeed.tech/tags/vs.md>)

### AI overview

Across 113 DeepSWE tasks, GPT-5.6 Luna had higher standalone quality than DeepSeek-V4 Flash 0731: pass@1 was 67.2% versus 53.3%, Luna led at every equal attempt count, and it was faster. DeepSeek cost about $0.10 per rollout versus Luna's $0.61 and yielded 532 solves per $100 versus 110. It also broke existing tests in a smaller share of its failures (9% versus 15%). A DeepSeek-first cascade that escalated rejected answers to Luna solved 78.9% of tasks at $0.385 each in this run, exceeding Luna alone's pass@1 while costing less. Luna led in 7 of 8 task domains and all 5 languages; DeepSeek led in query and config work, but performed especially poorly on JavaScript. These results are specific to the reported benchmark run.

### Source excerpt

We ran 900 DeepSWE rollouts on DeepSeek-V4 Flash and GPT-5.6 Luna. Luna leads pass@1 by 14 points; DeepSeek delivers 4.8x the solves per dollar.

## Kimi K3: the complete developer guide

DevFeed: [Kimi K3: the complete developer guide](<https://devfeed.tech/articles/kimi-k3-the-complete-developer-guide-80195.md>)

Original publisher: [Read original article](<https://www.together.ai/blog/kimi-k3-guide>)

Author: Zain Hasan, Shobhit Dixit

Published: 2026-08-01T00:00:00Z

Content type: tutorial

Language: en

Sources: [Together.ai](<https://devfeed.tech/sources/together-ai.md>)

Topics: [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Streaming](<https://devfeed.tech/topics/streaming.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-api](<https://devfeed.tech/tags/ai-api.md>), [api](<https://devfeed.tech/tags/api.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [code](<https://devfeed.tech/tags/code.md>), [costs](<https://devfeed.tech/tags/costs.md>), [developer-guide](<https://devfeed.tech/tags/developer-guide.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [examples](<https://devfeed.tech/tags/examples.md>), [guide](<https://devfeed.tech/tags/guide.md>), [kimi-k3](<https://devfeed.tech/tags/kimi-k3.md>)

### AI overview

This guide explains Kimi K3's architecture, API usage on Together AI, context handling, tool calling, pricing, and benchmark results. It covers practical integration details such as streaming, structured outputs, dynamic tool loading, and caching, alongside the model's open weights and reported performance.

### Source excerpt

Kimi K3 is the first open 3T-class model. See how it benchmarks, what it costs, and how to call it on the Together AI API, with copy-paste code examples.

[Next page](<https://devfeed.tech/sources/together-ai.md?cursor=WyIyMDI2LTA4LTAxVDAwOjAwOjAwKzAwOjAwIiwgIjY0ZGY5NzUzLTI2YjUtNGJjYy1hM2M4LTcyNmU0NmNjNzgyZiJd>)