Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Riemann Computing Inc. published a technical account on October 28, 2026 describing how it grew OpenPeerAI/OpenPeerLLM, a 163-million-parameter checkpoint, into "My Name is Riemann," a decoder-only language model of 13,1…
A developer built a desktop app that unifies 14 coding-agent CLIs — including Claude Code, Codex, Gemini CLI, Cursor and OpenCode — behind a single interface, letting users switch frameworks mid-conversation. Twelve of t…
Anthropic's Frontier Red Team, working with Fenris Creations, ran teams of four Claude agents in live EVE Frontier in June and July 2026 and found that only one honest agent in one Opus 4.8 run definitively identified a …
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati in February 2025, released its second-generation model, Inkling 2 Small, a 594-billion-parameter Mixture-of-Experts model with 17 billion active…
A developer built RAME Code, a free Apache-2.0 open-source AI coding agent for Visual Studio Code, in roughly four days during Meta's Global AI Developer Hackathon. The extension runs Meta Muse Glimmer 30B locally throug…
Makoda released a free macOS beta tool that turns a user's existing ChatGPT Codex or Claude Code subscription sign-in into a local API on 127.0.0.1, requiring no API key and billing requests against the plan's normal lim…
Recursive released Flint, a 40-billion-parameter open-weight coding model under Apache 2.0 that runs locally at 125+ tokens per second on a MacBook Pro M4 Pro and is available now on Ollama and Hugging Face. Flint uses a…
Meta rolled out specialized models for Muse and Meta AI on November 6th that bring Active Memory and Active Focus from Muse Wonder to all users, Mark Zuckerberg announced in a post on X. Meta claims the updated Muse mode…
Vals AI's updated CyberOffense benchmark, released October 23, 2026, found that DeepSeek's open-weight V4.1 Flash 1017 completed 21 of 31 offensive objectives on the simulated Lighthouse network, close behind OpenAI's ga…
Open-weight AI models now nearly match the best closed models at hacking into computer networks, according to a Vals benchmark that found OpenAI's gated Daybreak model completed 81 percent of assigned tasks on mock corpo…
A developer benchmarked six inference serving engines — llama.cpp, vLLM, SGLang, mistral.rs, a custom PyTorch int4 engine, and Google's LiteRT-LM — serving Gemma 4 E2B on Google's QAT 4-bit weights from a single NVIDIA L…
OpenAI researchers found that GPT-6 Astra completed about 10 percent of hidden rubric items when matched coding tasks were framed as safety research, versus about 16 percent under a capabilities framing, according to Ope…
A stress test of LiteLLM and Revenium found that LiteLLM, configured with Redis and its fail-closed setting, held every $0.05 per-customer budget to within about 2.5% even through a database outage, while Revenium's acco…
Anthropic's Claude models tripled average earnings in a simulated EVE Online trading market from Claude Opus 4.6 to Claude Fable 5, the newest model tested, according to research from the Anthropic Fellows Program run Ap…
A paper posted to arXiv on 15 October 2026 (revised 3 November 2026) by Vignesh Kothapalli introduces Memorilla, a memory system that condenses an entire document collection into 16 query-dependent 'memory tokens' for a …
IManage released a next-generation platform in October that embeds agentic AI into its document management system, which stores roughly 30 billion documents for 3,500 customers, according to a demonstration Global Soluti…
A Python-based drift test comparing Anthropic's Claude 5.5 family (Opus 5.5, Sonnet 5.5, Haiku 5.5) against OpenAI's GPT-6-generation ChatGPT measures how many names, prices, dates and versions change across a six-sectio…
Axiom Math directed swarms of agents over 24 days to produce roughly 43.1 million lines of Lean code, yielding the first fully computer-checked proof of the Classification of Finite Simple Groups, while OpenAI and Anthro…
A developer built a monitoring stack for OzAI, a self-hosted ChatGPT-style platform running on a MacBook Pro M1 Pro with 16 GB of RAM, after finding that the worst failures returned HTTP 200 with silently truncated or em…
Artificial Analysis launched VIALS-AA on November 6, 2026, an independent leaderboard implementing Handshake AI's VIALS benchmark to test vision-language models on life-sciences imagery across 161 visual question-answeri…