Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Enterprise teams including Shopify, Meta, Intercom, Ramp, Snowflake, and ONESTRUCTION have replaced frontier LLM API prompting with compact open models of 8B to 35B parameters trained via reinforcement learning against a…
Anthropic lowered Claude Code's default reasoning effort from high to medium in March, fixed a caching bug that stripped older reasoning from long sessions, and reverted a response-shortening system instruction that caus…
OpenAI's GPT-6 Astra requires noticeably less hand-holding than prior models like Opus, making fewer wrong assumptions about developer intent, according to hands-on testing described in a step-by-step guide to building a…
Anthropic's Claude Opus 5 posted top scores on Anthropic's own benchmarks — more than doubling Opus 4.8 on its frontier coding test and tripling the next best model on a separate agentic benchmark — but developers includ…
A week of head-to-head testing found GPT-6 Astra outperforming Fable 5.1 on most real-world coding and reasoning tasks despite roughly equal benchmark scores, according to developer testers cited in the comparison. Teste…
Token Rhythm's NeoHorse-1-4B, a 4 billion parameter model built on Qwen 3.5, was trained on live router decision logs rather than a static dataset, with a stronger teacher model correcting its attempts in real time and t…
Edge0-35B-A3B, a 35B-parameter sparse mixture-of-experts model built on Qwen3.5-MoE, loses an average of 3.9 points across five benchmarks when quantized to 4-bit through its edge0 inference pipeline versus its fp16 base…
Nex-AGI released its Nex-N2.5 family of open-weight agentic models in three sizes — mini, Pro, and Max — with a prebuilt Docker image running a customized SGLang fork (nexagi/sglang:v0.5.18-nex-patch) as the supported de…
Qwen released Qwen-Drive-1.0-4B, a 4-billion-parameter vision-language model built on Qwen3.5-4B that fuses perception and planning into a single model with a bird's-eye-view perception head and two planner variants, pub…
Anthropic, Google and OpenAI each shipped two versions of their best models between September 1 and 3, 2025 — a public tier priced at $10 per million input tokens and $50 per million output tokens for Claude Fable 5.1 an…
Simon Willison's Weblog published a post on September 12, 2026 titled "Generating running routes with GPT-6 Astra and ChatGPT Work," describing the use of GPT-6 Astra and ChatGPT Work to generate running routes. The sour…
Cognition released SWE-2, a coding model post-trained with reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3, scoring 50.0% on FrontierCode 1.1 Main — within 1 point of Fable 5.1 at 64% lower cost. SWE-2 h…
A developer catalogued a 1,245-object database schema with an LLM, generating natural-language descriptions for each table, and found that retrieval quality dropped sharply rather than improving. The token "contact" went…
A developer tutorial recommends wrapping large language model calls with Python and Pydantic schema validation plus an Instructor middleware layer to block prompt injection and malformed output, citing a three-day produc…
The open-source project local-deep-research reported approximately 95% on SimpleQA (n=500) and 77% on xbench-DeepSearch (n=100), which it says makes it the first fully-local open-source project to hit those marks on a si…
A developer built an hourly e-ink newspaper pipeline that pulls roughly 120 items from about 100 RSS feeds, uses Claude Haiku 4.5 on Bedrock as an automated editor, and typesets a 480x800 one-bit page for the Xteink X4 e…
A post titled "Why Beyond Traditional RAG?" linking to a Gumroad page for "Advanced RAG LangGraph Patterns" drew 7 points and 4 comments on Hacker News. The submission points readers to a paid resource on advanced retrie…
OpenRouter retired B3IT monitoring for the deepseek/deepseek-v4-flash-0731 endpoint on coreweave/fp8 on 2026-09-14 after failing to find enough border inputs, and its logprob probe also failed because the endpoint return…
Monitoring of the deepseek/deepseek-v4-flash-0731 endpoint at morph/bf16 was retired on 2026-09-14 after the B3IT border-input detector could not find enough border inputs, while the endpoint's logprob probe also failed …
Claude Desktop users connect Model Context Protocol (MCP) servers by editing the claude_desktop_config.json file, located at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\c…