Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Supermicro, IBM and NVIDIA announced a KV cache storage platform that cuts time to first token by 56x on 130,000-token prompts versus a GPU that recomputes the context from scratch, backed by a reference architecture IBM…
OpenAI introduced GPT-6 Cyber, an AI model built to find software vulnerabilities, develop exploits, and suggest patches. According to OpenAI, GPT-6 Cyber found 81 previously unknown vulnerabilities in Firefox and develo…
Google has added Gemini Omni 1.2 Flash to Google Vids, letting users generate video clips of 3 to 15 seconds with improved character consistency and vocal realism. The update adds a duration slider for pacing, better ref…
Tencent announced the release of Tencent Hy4, its latest-generation large language model, which the company says ranks among the world's leading open-source AI models, with model weights and system cards to follow in the…
A developer built a post-hoc rewrite-budget system for model-drafted documentation that measures a per-section edit-distance ratio between the model's draft and the text a human actually merged, then revokes drafting rig…
Moonshot AI published the full open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model previously available only through its API since October 23, and its own benchmark table puts the model at 65.7 on Terminal-Ben…
Anthropic is improving its AI models at roughly twice its prior rate, according to a Nov 5 independent assessment by nonprofit evaluator METR, which found Anthropic's AECI benchmark index has grown at 1.9 times its earli…
Researchers behind the Alpamayo-R1 autonomous driving project demonstrated a native vision-language-action (VLA) policy that applies safety rules at inference time through attention steering rather than retraining, compl…
Forum users on Level1Techs report that running local LLMs on a Proxmox host with an Nvidia RTX 3090 passed through to a VM works without known blockers, with one user saying their local agents run in a Proxmox VM with "s…
Mistral released the weights for Mistral Large 4, a 1.05 trillion-parameter mixture-of-experts model nicknamed "Le Chonk," on October 29, making it the largest open-weight model published by any European or American lab.…
METR reported on November 5, 2026 that Anthropic's internal Model 4 became the first model in its analysis to score above 50% on every benchmark it tracked, posting 178.6 on the AECI, Anthropic's variant of the Epoch Cap…
Anthropic's latest internal AI model, called Model 4 by METR, scored above 50% on every benchmark it was tested on, the first time any model in METR's dataset has cleared that threshold across all tested benchmarks, acco…
Preferred Networks released PLaMo 3 Translate on 5 November 2026, an open-weight family of translation-specialized LLMs in four sizes from 0.8B to 31B parameters, available on Hugging Face and covering 53 languages. Buil…
Google's TPUv7 (Ironwood) delivered 1.42 to 1.67 times better performance per dollar than Nvidia's GB300 NVL72 on the AgentX agentic inference benchmark, according to SemiAnalysis. The TPUv7 submissions ran the 397-billi…
METR has indicated that Anthropic's AI development acceleration may have reached up to three times its previous rate, with roughly 80% of that increase attributed to AI based on Anthropic's internal data and model perfor…
Charge Robotics cofounder and CTO Max Justicz used OpenAI's GPT-5.6 Sol model, accessed through the Daybreak Blue tier, to find CVE-2026-109518, a root-level man-in-the-middle flaw in APT, the package manager used by Deb…
A developer released THX-02, a 4.5B-parameter open decision model under Apache-2.0 weights, runtime and server, with a free hosted API. The model returns calibrated probability distributions over user-supplied options (c…
Kinference, an inference control plane for Kubernetes and vLLM, published benchmarks comparing its KV-cache-aware load balancing against a plain Kubernetes Service across two identical pools of three vLLM 0.31.0 pods run…
OpenAI announced ChatGPT Cyber, a cybersecurity-focused model available to the highest tier of its Daybreak Red offensive-security program, after the same model was used to scan the Linux kernel and prepare patches for 9…
Oracle introduced Fusion Claw on September 29, 2026, a governed agentic execution runtime for its Fusion Cloud Applications that pairs large language models with deterministic enterprise computation so AI agents can plan…