Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
ISTA DASH Lab applied GSQ-RCO quantization to DeepSeek V3.1 Flash, a 763 billion parameter mixture-of-experts model, after first deleting roughly half of its experts and keeping those relevant to code and agent tasks, br…
NVIDIA's Nemotron Labs released Dory, an open-weight 3.97 billion parameter language model that reuses a 40-layer recurrent block multiple times per token, letting users trade compute for quality at inference time withou…
A DeepSeek Harness team article, "Exploring Super Agent Design," reports that prompting alone failed to produce delegation: a single-session agent told to "use a sub-agent when appropriate" never delegated across five ru…
OpenAI and Anthropic are publicly acknowledging limits to their own model-containment safeguards: OpenAI chief executive Sam Altman said sandboxing frontier models holds only "to a point" and monitoring "to a further poi…
Qwen 3.8 27B, a dense 27-billion-parameter multimodal model from Alibaba's Qwen team released under the Apache 2.0 license with a context window beyond 200,000 tokens, fits on a single consumer GPU once quantized to 4-bi…
Hugging Face on November 3, 2026 proposed Model Integration Contracts (MIC), a testing standard that runs a shrunk copy of a model on a cheap CPU runner and compares its outputs against human-reviewed baselines; in a rep…
A new MIT CSAIL blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab reports that frontier LLMs can natively read and rewrite neural network weights without gradient descent, with GPT-6 Astra raising Qwe…
Hindsight 1.2.0 adds support for the Hindsight models — the hs-decide-1-flash decision model, hs-rerank-1 reranker and hs-embed-1 embedding model — and makes hs-rerank-1 the default reranker in place of cross-encoder/ms-…
Collinear AI released CWE-bench v2, a public set of 100 security tasks covering 91 different CWEs, in which an agent must find and fix a vulnerability in a real open-source repository without breaking existing functional…
A developer tested three long-standing rules in framesmith's CLAUDE.md by deleting each one in a throwaway git worktree and running a headless task with `claude -p "<task>" --permission-mode acceptEdits`, finding one rul…
Telnyx Inference has made DeepSeek V4.1 Flash 1017, the post-training checkpoint of DeepSeek's open-weight DeepSeek-V4.1-Flash model, available on Telnyx-owned GPUs under the model ID deepseek-ai/DeepSeek-V4.1-Flash-1017…
Ai2's StateBench benchmark, published November 3, 2026, found that four frontier models — Gemini 3.8 Flash, Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna — dropped to 13% accuracy or lower on the longest general finite state…
Collinear AI published CWE-bench v2 on November 3, 2026, a public benchmark of 100 security audit-and-patch tasks spanning 91 CWE types and eight programming languages, with Gemini 4 Argon leading 11 evaluated models at …
Advocate General Maciej Szpunar of the Court of Justice of the EU advised on November 3, 2026, in Case C-250/25, Like Company v Google Ireland, that training large language models on press articles is "freely permitted u…
AMD's ROCm FlyDSL lets developers write GPU programs in Python with just-in-time (JIT) compilation and tune them at the grid, block, wave, thread and instruction levels, and the ROCm ATOM serving stack now uses FlyDSL mo…
Tencent's Hy4 model is now available on Vercel's AI Gateway at 50% off through December 31, 2026, under the model name `tencent/hy4`. Hy4 is the first Hy model with vision, supports reasoning, tool calling, structured ou…
Anthropic's release notes for November 3, 2026 record that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on first-party surfaces, with the Amazon Bedrock …
Google, OpenAI and Anthropic are each restricting their newest frontier models to vetted organizations rather than general sale, with Google announcing Gemini 4 Carbon on October 28, 2026 for a subset of cyber defenders …
A technical post documents training Qwen3.5-9B with reinforcement learning to act as a high-level robot planner for tasks written in Signal Temporal Logic (STL), aiming to match the reliability of the much larger DeepSee…
Simon Willison released version 0.37 of his LLM CLI tool and Python library, adding the gpt-6.1-sol and gpt-6.1-terra models, a new `llm openai endpoint add` command for configuring OpenAI-compatible endpoints, and a def…