Mistral Large 4 tech report: 52B-active/1T MoE, fits on four GB300 GPUs at 606GB
aronchick · x · 2026-11-03
Mistral released the tech report for Mistral Large 4 (ML4), a multimodal mixture-of-experts model with 52B active and 1T total parameters, built for reasoning and long-horizon agentic tasks.
Key points:
- The team quantified the penalty KV-cache compression inflicts on downstream agentic tasks and fixed it via model surgery followed by sparse-attention adaptation
- A scalable agent-based pipeline authors and verifies executable environments for software engineering, cybersecurity, and knowledge work
- Benchmarks: 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, 50.9% on the Artificial Analysis Cyber Index
- Serving combines sparse attention, optimized context parallelism, prefill-decode disaggregation, and speculative decoding; calibrated NVFP4 quantization shrinks the checkpoint to 606GB while matching FP8 reference, enabling deployment on four GB300 GPUs
- Multiple-teacher distillation provides high/medium/low reasoning-effort modes to trade token consumption against performance
Related event: Mistral Releases Large 4 Open Weights: 1T-Parameter MoE Model(5 posts)→
More from coding & agent
- Phone Harness MCP lets any MCP agent control real cloud iPhones and Androids — Scobleizer · 2026-11-03
- LangChain on late-2026 context engineering: browser use, long-term goals, async work — hwchase17 · 2026-11-03
- MotherDuck launches Flight Explainer: live charts for agent-built Python pipelines — josh_wills · 2026-11-03
- Vibe coding's free ride is ending: one simple app racked up a $13,594 bill — michalmalewicz · 2026-11-03
- Swarms v17 'InterLink' launches, claiming 10-40x speedup over LangChain and AutoGen — KyeGomezB · 2026-11-03
- Solo dev builds cozy port sim game HarbourLine entirely with Claude Code and Opus 5.5 — Kyrannio · 2026-11-03