Mistral Large 4 goes open-weight: 1T-param MoE with 52B active, runs on four GB300 GPUs
_AndrewZhao · x · 2026-11-03
Mistral researcher Pierre Stock announced that Mistral Large 4 (ML4) is now open-weight, with a technical report detailing its development.
- Specs: a multimodal mixture-of-experts model with 52B active / 1T total parameters, built for reasoning and long-horizon agentic tasks
- Benchmarks: 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, 50.9% on the Artificial Analysis Cyber Index
- Training insight: the team quantified how KV-cache compression hurts downstream agentic tasks, fixing it via model surgery followed by sparse-attention adaptation; a scalable agent-based pipeline authored and verified executable environments for software engineering, cybersecurity, and knowledge work
- Serving: calibrated NVFP4 quantization shrinks the checkpoint to 606 GB while matching the FP8 reference, enabling deployment on four GB300 GPUs; multiple-teacher distillation offers high/medium/low reasoning-effort modes
Mistral says it is far more capable and compact than the Preview version released three weeks ago.
Related event: Mistral Releases Large 4 Open Weights: 1T-Parameter MoE Model(5 posts)→
More from Models
- Claude Opus 5.5 is the main model behind this comedy AI, creator reluctantly admits — ivan_bezdomny · 2026-11-03
- Mistral's ML4 report: KV-cache-saving attention tricks regressed SWE-bench agentic scores — _AndrewZhao · 2026-11-03
- WaPo Investigation: LLMs Got Voters' Congressional Districts Wrong 48% of the Time — Polymarket · 2026-11-03
- fchollet: LLM writing's worst habit is turning every sentence into a punchline — fchollet · 2026-11-03
- User on xAI's $1,200/yr Grok plan still hits 100% weekly usage cap — eter_inquirer · 2026-11-03
- DeepSeek's releases may be the year's most important: run a whole business on <$50/month — EXM7777 · 2026-11-03