• October 30, 2026
Mistral Large 4 Goes Fully Open-Weight: Trillion-Parameter Model Tops Open-Source Coding Charts in Europe’s Sovereign AI Bet

France’s Mistral has released the full weights of its flagship Mistral Large 4 (Le Chonk). After three extra weeks of reinforcement learning, the trillion-parameter multimodal model now tops open-weight coding benchmarks (FrontierCode 2.0 at 19.7%, Terminal-Bench 4.0 at 41%) and scores 51% on cyber defense, autonomously finding and patching vulnerabilities. Trained entirely in Europe and self-hostable, it anchors Mistral’s Sovereign AI strategy — though self-hosting demands ~1.06 TB of weights and serious compute, and its license restricts some hosted services.

  • October 20, 2026
CERN Teams Up With Mistral: Timestral Beats Google at 209M Parameters, and Europe’s First Collider-Physics AI Model Is Coming

Mistral fired twice in one day: Timestral, its first time-series foundation model family, tops all three major benchmarks with just 209M parameters — beating Google’s TimesFM-3 — and runs inference on CPUs without a GPU. The same day, Mistral signed a strategic partnership with CERN to build the first foundation model for collider physics, trained on billions of simulated collisions. Europe is betting on open weights plus fundamental science, and the barrier to enterprise predictive analytics just dropped.

  • October 16, 2026
Apple and Meta Bury the Hatchet: Muse Spark Open Puts an 816B-Parameter Open-Weight Model on the Mac

Meta has released Muse Spark Open, an 816-billion-parameter open-weight model for coding and AI agents, in a surprise partnership with Apple — the rival it was feuding with just two weeks ago. Optimized to run locally on two M5 Ultra Mac Studios with 1TB of unified memory, it is the most capable US open-weight model to date, aimed squarely at Chinese open-source rivals.

  • October 12, 2026
Nvidia in Talks to Acquire Reflection AI: Jensen Huang’s Open-Weight Empire Play May Use Acquihire to Dodge Regulatory Review

Nvidia is in early talks to acquire open-weight AI startup Reflection AI or deepen its investment, with a deal potentially landing within weeks, the Financial Times reports. Valued at $25 billion pre-money, Reflection has already taken $800 million from Nvidia. The FT highlights a striking twist: Nvidia may structure the deal as an “acquihire” — hiring staff and licensing tech without buying the company — specifically to sidestep a lengthy regulatory review. It’s the latest move in Jensen Huang’s open-weight empire building, following the $12.9B Hugging Face purchase and the $20B Groq licensing deal.

  • October 8, 2026
Mistral’s Le Chonk: Europe’s First Trillion-Parameter Open-Weight Model Beats Closed Giants on Cybersecurity

French AI firm Mistral has launched Le Chonk (Mistral Large 4), a 1-trillion-parameter open-weight model trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers. It scores a world-best 82% on vulnerability reproduction, while Claude Opus 5.5 and GPT-6 Astra score near zero for refusing the task. Weights arrive end of October.

  • August 17, 2026
Z.ai Releases GLM-5.3: Open-Weight Model Outscores Anthropic and OpenAI on CyberGym

Chinese AI company Z.ai has released GLM-5.3, an open-weight model that scored 84.5% on the CyberGym cybersecurity benchmark, beating Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol. Built on the same base as GLM-5.2 with all gains from scaled post-training, it also set an open-source record on Terminal Bench 3.0 and unexpectedly acquired exploit-chain reasoning, uncovering 1,097 critical vulnerabilities in Linux, WebKit and FreeBSD. Weights open within two weeks.