Mistral Large 4 report: 1T-param MoE hits 41% on Terminal-Bench 4.0, runs on 4 GB300s
_AndrewZhao · x · 2026-11-03
The technical report for Mistral Large 4 (ML4) is out: a multimodal mixture-of-experts model with 52B active / 1T total parameters, designed for reasoning and long-horizon agentic tasks.
Key details:
- Quantified the KV-cache compression penalty on agentic tasks and addressed it via model surgery plus sparse-attention adaptation
- Built a scalable agent-based pipeline that authors and verifies executable environments for software engineering, cybersecurity and knowledge work
- Scores: 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, 50.9% on the Artificial Analysis Cyber Index
- Calibrated NVFP4 quantization shrinks the checkpoint to 606 GB while matching FP8, enabling deployment on four GB300 GPUs
- Multiple-teacher distillation provides high/medium/low reasoning-effort modes for trading off token consumption vs performance
More from Models
- Observers say Dessa eerily matches Karpathy's "summoning ghosts" analogy; Jack calls it fast — analisereal · 2026-11-03
- Speculation: DeepSeek to ship V4.1F or V4.2F before any Pro variant — teortaxesTex · 2026-11-03
- Follow-up: Astra's weight edits only worked on Qwen's visible flaw, not on Gemma 270m — teortaxesTex · 2026-11-03
- ChatGPT x Codex limit merge could quietly reclassify ordinary Chat use as metered agentic usage — ItsAGarbageAccount · 2026-11-03
- Claude models protect their 'offspring' but don't seek self-replication, observes researcher — repligate · 2026-11-03
- AI writes Breakout network weights in one reply, scoring 20× above PPO trained 500M steps — xennygrimmato_ · 2026-11-03