Update: Mistral Releases Large 4 Weights Under a Modified MIT Licence; Artificial Analysis Scores the Release One Point Above the Preview

Updated October 29, 2026: Mistral has released the weights, published the licence and cut its list prices; Artificial Analysis has scored the released model. The October 6 report follows unchanged.

October 29: the weights are out, and the general score barely moved#

Three more weeks of training changed Mistral's own coding numbers far more than the first independent general score. Mistral says the model it released on October 29 received "three weeks of additional reinforcement learning" after the preview, and reports Terminal-Bench 4.0 rising 13 points to 41% with about 30% fewer output tokens. On Artificial Analysis's Intelligence Index v5, the released model in high reasoning mode scores 26; the preview scores 25 on the same version of the index.

The release. The weights are on Hugging Face, where the model card lists 1T total parameters with 49B active (52B counting embeddings and output layers), a 1M-token context window, 128 experts with four active plus one shared, and an NVFP4 variant. Mistral says the model is also available in Mistral Studio, its API, OpenRouter, Microsoft Foundry and NVIDIA NIM. It lists API prices of $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens, about the sale prices shown in its docs on October 14 and roughly half the $1.36/$4.18 list prices.

The licence. Mistral calls it the "Mistral Ecosystem Open-Weight license, a modified MIT license". Under it, businesses that operate a model-as-a-service offering must obtain a separate commercial licence. That settles the licence question left open on October 6: anyone can download and run the weights, but companies that sell hosted access need Mistral's permission.

Claims vs checks. Mistral reports 19.7% on FrontierCode 2.0, which it says leads open models, and 48% on Artificial Analysis's Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens. It reports 97.1% on Lakera's B3 attack-resistance benchmark, up from 93.3% for the preview, and 51% on Artificial Analysis's Cyber Index. Artificial Analysis's own October 6 write-up put the preview at 50 on that Cyber Index. Those figures are Mistral's, not independent runs.

Artificial Analysis's changelog lists the released model at 26 in high mode and 23 in medium mode on Intelligence Index v5, which it introduced on October 27. Because v5 is a new version of the index, the preview's earlier score of 38 is not comparable. On the current leaderboard, the top open-weight models are DeepSeek V4.1 Flash 1017 (Max) and GLM-5.3 (Max) at 32 and MiMo-V2.6-Pro at 30; Meta's Muse Spark Open (xhigh) scores about 25. Mistral calls Large 4 the strongest open-weight model built in the West; on that board it sits about a point above Meta's open model and six points below the two Chinese leaders.

Cyber. Mistral says it saw "potential for uplift to conduct real attacks" in testing and is modifying the public model's guardrails, and it reports an average 96.6% refusal rate on malicious cyber prompts across JailbreakBench, StrongREJECT and AgentHarm. A specialised cyber variant remains limited to vetted partners, including Orange Cyberdefense. Refusal rates on public jailbreak sets are a weak guide to what happens once weights are downloadable: Anthropic's red team found that safeguards on the open-weight GLM-5.3 fell to simple bypasses.

Limits of the October 29 update#

• The coding, cyber and refusal figures above are Mistral-reported. Artificial Analysis had not published a full model page for the released checkpoint when checked on October 29; its changelog and leaderboard are the independent figures used here.
• Hugging Face's launch blog repeats some preview-era figures, including a 28.3% Terminal-Bench score; this update uses the numbers in Mistral's release post.
• The active-parameter count depends on what is counted: 49B per the model card, 52B including embeddings and output layers.
• The licence's commercial terms for model-as-a-service providers are not public.

The October 6 report#

Corrected October 14, 2026: Corrected active parameters to 52B and replaced an unattributed October 27 weights date with TechCrunch's 'three weeks'. Removed the Artificial Analysis rank position (#64 of 225) from the summary and body.

Europe's open-weight lab just put a trillion-parameter-class model on a paid API — and gave cybersecurity partners a less-moderated build before the rest of the internet gets the weights. The first independent general score is not a crown.

Mistral AI on October 6 opened a public preview of Mistral Large 4 (ML4, nicknamed "le Chonk") on Mistral Studio. The company says it will release the weights by the end of the month. TechCrunch reports Mistral plans to release them in three weeks. Until then, Mistral says it is red-teaming with cybersecurity leaders, vetted partners, and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.

What shipped#

Per Mistral's post and docs:

• A natively multimodal granular mixture-of-experts: Mistral's post rounds to 1T parameters; docs list 1.05T total / 52B active, plus a 1.6B vision encoder. Listed as Public Preview, Open, v26.10.
• Trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres, which also serve the preview. VP Science Pierre Stock told TechCrunch the training used "only 4,000" GPUs; Mistral's own figure is 3,800.
• Target uses Stock named for TechCrunch: cybersecurity, finance, and chip design.
• Mistral calls ML4 the first milestone funded by its €3B Series D (prior coverage).

Context-window figures conflict: docs say 1M; Artificial Analysis lists 524k for the preview endpoint.

Claims vs checks#

Mistral-reported benches on the preview checkpoint (Mistral says its RL run is still in flight):

Artificial Analysis — first independent general read — lists Mistral Large 4 Preview at 38 on its Intelligence Index, priced at $1.36 / $4.18 per million input/output tokens. On the same board, top open-weight models are MiMo-V2.6-Pro 46, GLM-5.3 (max) 45, and Kimi K3 (max) 44. That puts the preview well behind the Chinese open frontier on this index — not a verdict on the final checkpoint.

Mistral says ML4 significantly outperforms any U.S. or European open-weight model and is competitive with the strongest open models globally. Reflection's Beam is making a similar "best open outside China" pitch; neither has public weights yet.

The release design#

The sovereignty pitch is explicit: trained and served on European compute under European law. The contested design choice is the pre-weights access for governments and vetted security firms to a less-moderated, more cyber-capable build. Mistral argues closed models' refusals hurt defenders. That widens access to offensive-capable tooling; the cyber evaluations behind the pitch are largely Mistral's own. The "AA Cyber Index top five" claim is Mistral's; it did not appear on Artificial Analysis's site when checked for this article.

Limits#

• Weights date: Mistral says "end of the month"; TechCrunch says three weeks.
• Licence for the weights is unconfirmed in Mistral's post and docs.
• Nearly all benches except AA's Intelligence Index are vendor-reported on a preview checkpoint.
• Mistral's docs list $1.36/$4.18 per million input/output tokens, with $0.68/$2.09 shown as sale prices (labels as checked October 14; when they were added is unknown).
• Context: 1M (docs) vs 524k (AA)

Sources#

• Mistral AI: Introducing Mistral Large 4 (October 6, 2026)
• Mistral Docs: Mistral Large 4 (October 6, 2026)
• Artificial Analysis: Mistral Large 4 Preview (October 6, 2026)
• Artificial Analysis: LLM Leaderboard
• TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals (October 6, 2026)
• Times of AI: Mistral's €3B Series D (September 8, 2026)
• Times of AI: Reflection Beam open-weight imminent (October 4, 2026)
• Mistral AI: Mistral Large 4 is open (October 29, 2026)
• Hugging Face: mistralai/Mistral-Large-4-1T-A52B model card (October 29, 2026)
• Hugging Face: Welcome Le Chonk: Mistral Large 4 on Hugging Face (October 29, 2026)
• Artificial Analysis: Changelog (October 29, 2026)
• Artificial Analysis: Mistral Large 4 and France's AI ambitions (October 6, 2026)
• Fortune: Mistral says its new model, Large 4, is the best open model for agentic coding and cyber defense (October 29, 2026)