Mistral Large 4 tech report: 1T-param MoE, 606GB quantized, runs on four GB300 GPUs

_AndrewZhao · x · 2026-11-03

Mistral released the technical report for Mistral Large 4 (ML4): a 52B-active / 1T-total multimodal MoE for long-horizon agentic tasks. Key points:

Quantified KV-cache compression penalties on agentic tasks, fixed via model surgery plus sparse-attention adaptation;
Scores 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, 50.9% on Artificial Analysis Cyber Index;
Serving combines sparse attention, context parallelism, prefill–decode disaggregation, and speculative decoding;
Calibrated NVFP4 quantization shrinks the checkpoint to 606GB while matching FP8, enabling deployment on four GB300 GPUs;
Multi-teacher distillation offers high/medium/low reasoning-effort modes.

The model is now open-weight, confirmed by Mistral researcher Pierre Stock.

More from Infra