Mistral's largest model, launched as a public preview on October 6, 2026 — ML4, or as Mistral insists, le Chonk. A 1 trillion-parameter natively multimodal MoE with 49B active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, which also serve the preview. Weights were promised under open weights by the end of October; until then it was API-only on Mistral Studio while the model was red-teamed with cybersecurity leaders, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities. Artificial Analysis scored the preview at 38 on its Intelligence Index v4.3.2 (64th of 225, above the 26 median) and at 25 on v5.0 after the October 27 recalibration. A large share of training data was multilingual across more than 160 languages, including every official EU language.

The weights shipped on October 29, as Mistral-Large-4-1T-A52B in FP8 plus an NVFP4 build, after three more weeks of RL that Mistral says mainly improved coding, cyber defense and token efficiency. The license is the new Mistral Ecosystem Open-Weight License, a modified MIT license that allows commercial use but requires a separate commercial license to run the model as a Model-as-a-Service business. The card gives 49B active parameters per token, or 52B counting embeddings and output layers, with 128 routed experts (4 active) plus one shared expert, a 1M context, image input, and low, medium and high reasoning effort. Artificial Analysis now lists it as open weights with a 1M context and scores the released model 26 on v5.0 at high effort (23 at medium). Mistral calls that the highest of any Western open-weights model. On Cognition's own FrontierCode 2.0 runs Mistral reports 19.7%, ahead of the other open models in its chart.

The technical report (30 pages, posted to the HuggingFace repo on November 2) is unusually candid about a wrong turn. ML4 was pretrained on 3,840 GB200 GPUs near Paris for 45 days (about 4.15M GPU-hours at 36% MFU) with heavy KV-cache compression: sliding-window attention in three of every four layers, cross-layer KV sharing and YOCO, which together cut the KV cache 16×. Perplexity and knowledge benchmarks looked fine, but SFT cooldowns on intermediate checkpoints underperformed on long-context agentic tasks such as SWE-bench and BrowseComp. Mistral then performed model surgery before the second pretraining phase. It removed the sliding windows and cross-layer sharing, copying KV weights into the affected layers and adding LeRoPE and SSMax, but kept YOCO. It then adapted the model to DeepSeek Sparse Attention over about 213B tokens. The report concludes that comparable language-modeling loss does not imply comparable long-horizon performance. Post-training uses agents that author and verify executable environments for software, terminal, cyber and knowledge-work tasks. The RL run generated 5.36M rollouts and 162.5B output tokens, with a CISPO-style loss under Muon. Low, medium and high reasoning effort come from distillation from three teachers. One aside: most of Mistral's BrowseComp search trajectories hit "query-reflection" pages left behind by earlier agents' searches by October. NVFP4 quantization shrinks the checkpoint to 606 GB, enough to serve on four GB300s, and an embedded DSpark draft head speeds decoding. Report figures include 41% on Terminal-Bench 4.0 and 50.9% on the AA Cyber Index.

The preview announcement's sharpest claim is about cybersecurity, and it is a claim about refusals as much as capability. On the AA Cyber Index, ML4 ranks top five globally and leads non-Chinese open-weight models; on the index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scores 82%, the highest of any model, and it solves 93% of Cybench's 40 competition exercises. Mistral's point is that Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse, and that defending software starts with proving a flaw is real — so provider-level refusals can block legitimate vulnerability research mid-incident. The counterweight is in their own safety section: ML4's average refusal rate on cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is reported as higher than every other OSS model they measured. It resists 93.3% of attacks on Lakera's public B3 benchmark and scores 1.691 of a possible 2 on KORABench.

Elsewhere the numbers are strong but the comparisons are narrower than the framing. Coding: 61.7% DeepSWE v1.1, 59.4% SWE-Atlas-QnA, 28.3% Terminal-Bench 4, a combined Coding Agent Index of 49.8% ahead of DeepSeek V4 Pro 0813 and Qwen3.8-Max — all taken from Artificial Analysis's coding index rather than self-run. Agentic: 59.9% on AutomationBench's 657 business workflows and 1,393 Elo on AA-Briefcase. Multimodal is where Mistral claims to pass closed frontier models, on visual grounding, by one point — Dense 200 42% against GPT-6-Astra's 41%. It reports state-of-the-art SciCode-Verified among open-weight models, and third-party vals.ai evaluations beating GPT-6-Astra on legal and financial tasks. Two honest results sit alongside those: a blind five-model human evaluation run with Surge AI put ML4 second (3.74), ahead of GLM-5.3 (3.60) and Kimi K3 (3.59) but well behind Claude Opus 5 at 4.22; and Mistral's own expert annotators preferred ML4 over GLM-5.3 on CAD and STEM while rating it only on par or close in finance and coding. The RL run behind the preview is described as still in flight and showing no saturation, producing roughly 33B tokens a day at 3K GPUs (about 16B trainable after filtering), funded by the €3B Series D Mistral calls the largest equity round ever raised by a European technology company.

Model Details

Architecture MOE
Parameters 1T
Active params 52B
Experts 128 (top-4)
Context window 1,000,000
AA Intelligence 26 was 25 on v5.0
Training hardware 3,840 NVIDIA GB200
License Mistral Ecosystem Open-Weight License (modified MIT)

Benchmark Scores

Benchmark Score Mode
Cybench 93% —
AA Cyber Index (reproduce-and-patch test) 82% —
DeepSWE v1.1 61.7% —
SWE-Atlas-QnA 59.4% —
Terminal-Bench 4 28.3% —
Coding Agent Index (combined) 49.8% —
AutomationBench 59.9% —
Lakera B3 (attack resistance) 93.3% —

Variants

Name Parameters Notes
Mistral-Large-4-1T-A52B — 2026-10-29 open-weight release, FP8; AA v5.0 26 (high), 23 (medium)
Mistral-Large-4-1T-A52B-NVFP4 — Post-training NVFP4 quantization, 606 GB, serves on 4x GB300

Paper

moemultimodalopen-weightagenticcodecybersecuritymultilingual

Related