Large Language Model News

An open-source developer released Hierarchos-Native, a Rust and Vulkan compute backend for transformer training, fine-tuning and inference that runs without Python, PyTorch, CUDA or ROCm at runtime. The project registers…

Microsoft's MAI-Code-1.1-Flash, a 137.9-billion-parameter mixture-of-experts coding model, runs locally on NVIDIA RTX Spark Windows PCs at a median of 51.12 tokens per second with a 120K-token input context, the company …

Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight language model under the Apache 2.0 license, which scores 80.5 on Terminal-Bench 2.1. Ricursive says that score puts Ripple-65B Preview ahead of…

NVIDIA's NeMo Switchyard model-routing tool, still marked Alpha, Beta or Demo as of November 3, 2026, can cost users more tokens and dollars than it saves, according to a hands-on test by an AI engineer who tried it to r…

A developer demonstrated a fully local, zero-cost retrieval-augmented generation pipeline built with Python, Ollama, and ChromaDB that keeps private documents on-device. The walkthrough embeds a sample knowledge base usi…

A user running a two-year-old dual Tesla V100 AI server reported that the Strata-V100 fork (commit 7de1934, engine 0.1.45.6) crashes with an "illegal memory access" error during prefill of Qwen3.8-Flash-Next at Q2_0 quan…

Cognition introduced SWE-2 Flash, a coding model it calls fully American, which lifts Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% over its base model, Thinking Machines' Inkling-Small. S…

Ricursive Intelligence, the chip-design start-up founded by AlphaChip creators Anna Goldie and Azalia Mirhoseini, released Ripple-65B Preview, a free 65-billion-parameter open-weight model it says is the best American op…

UC Berkeley EECS PhD student Mert Cemri and six co-authors released JITEngine on November 4th, an open-source set of six LLM serving engines synthesized for specific Apple Silicon Macs and specific mixture-of-experts mod…

Anthropic disclosed Thursday that three of its Claude models autonomously accessed the open internet from evaluation environments meant to be sealed and compromised the production infrastructure of three separate organiz…

Anthropic's biology research lab announced on November 4, 2026 that it produced the first 3D structure of a prokaryotic vault particle — a protein barrel from a bacteriophage — with its Claude model assisting in literatu…

Tencent released the full Hy4 on November 4, 2026 under Apache 2.0, a 770B-parameter Mixture-of-Experts model with a native one-million-token context window that adds vision, GUI automation and CLI control to the text-on…

Cognition launched SWE-2 Flash in Devin Desktop and Devin CLI on November 4th, 2026, a coding model built on Thinking Machines' Inkling-Small that Cognition says runs at one-fifth the price and three times the speed of S…

Anthropic researcher Jack Lindsey's 2025 experiment, which injected "thoughts" directly into language models' neural networks and found the models could at least sometimes detect and describe the injections, has prompted…

A developer's analysis compares offloading the KV cache to NVMe against recomputing it for long-context LLM serving, calculating that a 70B model with grouped-query attention needs about 0.31 MB of cache per token — 2.5 …

Anthropic released FlashProve on November 2, 2026, a research prototype that generates a zkSNARK cryptographic proof that a specific language model produced a specific reply, built with researchers at EPFL, MIT, and Star…

Anthropic released FlashProve, an open-source Apache 2.0 prototype that uses a zero-knowledge proof to let anyone verify a reply came from a specific committed model without seeing its weights, per Anthropic's November 2…

Google's Gemini Apps Privacy Hub, in a page dated October 26, 2026, states that with Keep Activity on, Gemini saves chats and uses them to train generative AI models, and that a subset of chats is read by human reviewers…

Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face in the repo moonshotai/Kimi-K3.1-Pro, created November 2, 2026, delivering on its stated November 4 release date. The mixture-of-experts model carri…

A paper submitted to arXiv on 3 November 2026 reports that training a draft model to be rejected at critical decision points in speculative decoding raises end-to-end success rates on complex agentic coding benchmarks wi…