# D1 Published articles for D1. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Zero-shot evaluation of open vision-language models as System 1 decision models on radiology tasks DevFeed: [Zero-shot evaluation of open vision-language models as System 1 decision models on radiology tasks]() Original publisher: [Read original article]() Author: Bradley J Erickson, Sara Salehi Published: 2026-11-03T05:00:00Z Content type: news Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Artificial Intelligence](), [smolvlm](), [Large Language Model](), [qwen]() Tags: [3d](), [compute](), [cs-ai](), [cs-cv](), [d1](), [decision-models](), [detection](), [devices](), [energy](), [evaluation](), [faster](), [images](), [language-models](), [llms](), [motion](), [multimodal](), [performance](), [qwen](), [vision](), [vision-language-models](), [zero](), [zero-shot]() ### AI overview In a zero-shot comparison across seven radiology tasks, Clef 27B generally matched or exceeded Qwen 3.8-27B, while d1 3B was close on report text but ranked below Clef on six of seven tasks. All three models extracted findings from reports with AUROC above 0.95, but none performed well at grading motion artifact severity. Clef beat its Qwen backbone on CT volume findings and line/tube positioning, and identified 90 of 100 FLAIR MRI volumes where Qwen chose none. Clef was faster per request on tasks with many questions per input; d1 used three to nine times less energy per request than Clef. The authors report that a single-answer-token LLM can have similar cost and performance to System 1 models, with a 3B model approaching 27B models on text but not images. ### Source excerpt arXiv:2611.01631v1 Announce Type: cross Abstract: Purpose: We compared two multimodal System 1 decision models, Clef 27B and d1 3B, with a traditional LLM (Qwen 3.8-27B) on seven radiology tasks covering text, radiographs, CT and MRI. Methods: Three models were compared: Clef, Qwen3.8-27B and d1. Three tasks were interpretation of text and 4 were detection of finding in images. The tasks were: 1. finding presence of finding in CT-RATE chest CT reports; 2. Presence of finding in PadChest chest radiograph reports; 3. the same 18 findings on CT-RATE chest CT but 3D volumes; 4. MR safety status of devices from AccessGUDID device records; 5. line and tube position on RANZCR CLiP chest radiographs; 6. motion-artifact severity on MR-ART T1-weighted brain MRI and 7. sequence identification on BraTS 2021 brain MRI. The primary endpoint was AUROC. There was no fine-tuning of the models on the tasks. Results: All models completed all tests without a failed request. All retrieved findings from reports almost perfectly (AUROC >0.95). Clef ranked higher than its Qwen backbone on CT-RATE volumes and RANZCR lower on MR-ART. The two did not differ on BraTS, GUDID or PadChest, and differed by 0.002 on CT-RATE reports. On BraTS, Clef correctly identified 90 of 100 FLAIR volumes while Qwen never chose it. d1 ranked lower than Clef on six of seven tasks. No model performed well on motion severity. Clef was faster per request than Qwen on tasks with many questions per input. d1 used 3 to 9 times less energy per request than Clef. Conclusions: System 1 models can perform important tasks in radiology with high accuracy, at lower compute cost than LLMs used in the usual way. When an LLM is constrained to a single answer token, its cost and performance is similar to System 1 models unless multiple questions are asked. A 3B System 1 model was close to the 27B models on report text but not on images. No model performed well on motion severity. ## Send BLE air quality data to Cloudflare D1 with Python DevFeed: [Send BLE air quality data to Cloudflare D1 with Python]() Original publisher: [Read original article]() Author: Anne Barela Published: 2026-10-29T21:19:06Z Content type: tutorial Language: en Sources: [Adafruit Industries]() Topics: [Cloudflare Workers](), [Bluetooth LE](), [Python](), [data-engineering]() Tags: [air-quality](), [ble](), [bluetooth](), [bluetooth-low-energy](), [cloudflare](), [d1](), [database](), [python](), [sensors](), [serverless](), [serverless-database](), [software](), [tutorial]() ### AI overview A HibouAir sensor broadcasts air-quality readings over BLE, and a USB-connected BleuIO dongle scans for its board ID and forwards the advertising data over serial to Python. The script decodes CO2, temperature, and humidity, prints them, and POSTs JSON to a Cloudflare Worker. The Worker checks an API key in the request header before storing readings in Cloudflare D1 (SQLite on Cloudflare's edge); a second endpoint returns the latest measurement. The excerpt says Python needs pyserial and requests, and BleuIO uses AT commands rather than the operating system's Bluetooth stack. Scan interval, serial port, board ID, Worker URL, and API key are kept in a configuration file excluded from version control. ### Source excerpt sheikh shuhad has posted a tutorial on Hackster.io that moves Bluetooth Low Energy sensor readings into a serverless database. A HibouAir sensor advertises air quality data over BLE. A BleuIO dongle plugged into a computer's USB port scans for that specific board ID and passes the advertising data to a Python script over the serial [...] ## OpenAI makes its Jev competitor widely available as decision model competition heats up DevFeed: [OpenAI makes its Jev competitor widely available as decision model competition heats up]() Original publisher: [Read original article]() Author: Tim Fernholz Published: 2026-10-22T19:00:00Z Content type: news Language: en Sources: [TechCrunch | AI News & Artificial Intelligence]() Topics: [Artificial Intelligence](), [OpenAI](), [LLM optimization]() Tags: [ai](), [api](), [benchmarks](), [calibrated-decision-model](), [classification](), [competition](), [cto](), [d1](), [datadog](), [decision-model](), [decision-models](), [decisions-api](), [deep-learning](), [developer](), [developers](), [exclusive](), [game](), [image-classification](), [images](), [jev](), [openai]() ### AI overview OpenAI's Decisions API makes choices in 100-180 milliseconds by reconfiguring ChatGPT Luna for faster, lower-cost inference, rather than using a specially calibrated decision model. The API is now available to all customers and adds voice and visual modalities; OpenAI also geographically targeted inference resources to reduce latency across regions. Speak reported cutting the time between actions from 1.4 seconds to 0.13 seconds, and other customer trials used it for higher throughput. Benchmarks vary: TypeSafe's latest Jev ranks seventh and third on two cited indices, while OpenAI's model ranks eighth and fourteenth; Jev leads the PacMan-based JevMan benchmark. The excerpt says the best balance of intelligence, cost, calibration and speed depends on the workload, and TypeSafe's CEO recommends trying models in users' own work. OpenAI says Luna is strong at image classification and calibration, while further latency reductions could enable per-frame video calls. ### Source excerpt OpenAI is launching its decision model to the world. Can it compete with the original Jev? ## 🍔🧠 How D1 Manages Millions Of Databases At The Edge DevFeed: [🍔🧠 How D1 Manages Millions Of Databases At The Edge]() Original publisher: [Read original article]() Author: Alexandre Zajac Published: 2026-10-19T15:30:38Z Content type: article Language: en Sources: [Hungry Minds]() Topics: [control-plane](), [Cloud Architecture](), [Cloudflare Workers](), [database-optimization]() Tags: [architecture](), [cloudflare](), [control](), [control-plane](), [d1](), [databases](), [edge](), [visibility]() ### AI overview Cloudflare's D1 is a serverless SQLite database designed to support millions of small databases. Its architecture keeps coordination off the query path: routing metadata is cached in Workers KV, while control-plane operations use distributed creation workflows, heartbeats, and pooled Postgres access. The article describes how separating data-plane, control-plane, and background work allows queries to avoid a central Postgres round trip while tolerating eventual consistency elsewhere. ### Source excerpt PLUS: Anthropic redesigns source control 🔧, MicroVM pause-resume architecture ⚡, Git hooks visibility 👁 ## Liquid AI's d1-ultra now available on AI Gateway DevFeed: [Liquid AI's d1-ultra now available on AI Gateway]() Original publisher: [Read original article]() Author: Zachary Chen Published: 2026-10-15T00:00:00Z Content type: release Language: en Sources: [Vercel News]() Topics: [AI Chat](), [context window](), [litellm](), [Claude]() Tags: [ai-gateway](), [ai-sdk](), [api](), [d1](), [decisions-api](), [human-review](), [inference](), [liquid](), [liquid-ai](), [scale](), [sdk](), [types]() ### AI overview Vercel's AI Gateway has added Liquid AI's d1-ultra, a multimodal model for classification, routing, and scoring over text and images. It returns structured answers with probabilities, supports multiple questions per request and a 32,768-token context window, and charges for input tokens only. The release also describes API integrations, logging, budgets, and gateway pricing. ### Source excerpt d1-ultra from Liquid AI is now available on AI Gateway. It is a multimodal decision model for classification, routing, and scoring across text and images, including visual defect inspection and checking damage claims against customer photos. Rather than generating text, d1-ultra returns structured answers with probabilities. Ask a Boolean question to get the probability of a yes/no answer, use Choice to select among named categories, or use Score to rate an ordered scale. You can ask multiple questions about the same input in one request and use the probabilities to decide when to automate an action or request human review. d1-ultra is designed for harder real-world decisions, with an emphasis on calibrated probabilities. It supports a 32,768-token context window and is billed on input tokens only, including image input, with no output-token charges. Use liquid/d1-ultra as the model name. The examples below ask the same question through the AI SDK, OpenAI-compatible Decisions API, and TypeSafe-compatible API: See the decision models documentation for question types and additional usage details. Try d1-ultra in the model playground. AI Gateway provides a unified API for accessing models and tracking usage and cost. Decision calls appear in logs and count toward budgets alongside other model requests. AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including on Bring Your Own Key requests. Browse all decision models available on AI Gateway. Read more ## Under the hood, off the hot path: How D1 scales to millions of databases DevFeed: [Under the hood, off the hot path: How D1 scales to millions of databases]() Original publisher: [Read original article]() Author: Lambros Petrou Published: 2026-10-13T16:20:18Z Content type: article Language: en Sources: [Cloudflare Blog]() Topics: [control-plane](), [Supabase](), [database-optimization](), [Service fabric]() Tags: [agent](), [agentic](), [caching](), [control-plane](), [d1](), [databases](), [deep-dive](), [developer-platform](), [edge-database](), [scale](), [sql](), [workers]() ### AI overview Cloudflare describes how it scaled D1's control plane to support millions of managed SQLite databases while keeping query traffic off centralized components. The article explains edge routing, Workers KV and Hyperdrive caching, self-correcting database heartbeats, asynchronous usage processing, and pre-aggregation to reduce database load and write contention. ### Source excerpt A deep dive into how D1 architected its control plane to scale and manage millions of databases, while keeping data plane query operations at the edge. We share guiding principles on how to build on the Developer Platform effectively.