Mistral Releases Large 4 Weights, but Hosts Need a Licence

Mistral AI said on 29 October 2026 that it is releasing the weights of Mistral Large 4, the 1-trillion-parameter model it calls Le Chonk, three weeks after a public preview. Its blog post says: “Today, we’re releasing the weights of this improved Mistral Large 4”. Reuters headlined the release as aimed at cybersecurity users.

Update, 3 Nov: Mistral’s Large 4 technical report. Mistral added a 30-page technical report to the model’s Hugging Face repository on 2 November, and the model card links to it. It says early versions of the model cut the KV cache, the memory a model keeps for its context, “by a factor of 16 compared to a full-attention baseline”, using sliding-window attention over 512 tokens plus two cross-layer sharing schemes, CLA and YOCO. Perplexity and knowledge benchmarks looked as expected, but SWE-bench Verified scores “decreased as we moved from full attention to SWA, and regressed further when applying CLA”. The report concludes that “models that perform equally well on language-modeling loss can nevertheless perform very differently on downstream agentic tasks”, and that longer training can mitigate the penalty. It shows the SWE-bench declines in charts rather than as numbers in the text.

Mistral says it then performed “model surgery” to make the model close to full attention, keeping only the YOCO sharing, and adapted it to DeepSeek Sparse Attention, which selects the top 2,048 tokens for each query. The report states pretraining took 45 days on 3,840 Nvidia GB200 GPUs near Paris, about 4.15 million GPU-hours at 36% model FLOPs utilisation, and that a 606 GB NVFP4 checkpoint runs on four GB300 GPUs. In Mistral’s account, the memory cuts that make serving cheaper cost the model agent quality.

The weights come under a new Mistral Ecosystem Open-Weight License, which the model card describes as “a modified version of the MIT license” for commercial and non-commercial use, with one carve-out: “operating a Model-as-a-Service business requires a separate commercial license with Mistral”.

Short explainer · Mistral Large 4 weights

A vertical news Short narrated from this piece’s headline and opening line.

Business Pill · Sovereign AI

An evergreen Business Engineer explainer on sovereign AI. It teaches the general idea only and does not discuss Mistral or Mistral Large 4.

The key insight: By our arithmetic from Mistral’s own figures, three weeks of extra training moved Mistral Large 4 from 28% to 41% on Terminal-Bench 4.0, a 46% relative gain, with 30% fewer output tokens, and the weights now ship under a licence that leaves hosted-service use to a separate deal.

What Mistral Released

The Hugging Face model card describes “a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers)”, “trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe”. It lists a 1M context window and three reasoning efforts, low, medium and high.

When Mistral opened the preview on 6 October, its announcement said “Weights drop end of this month”, and its documentation listed weights and licence as “Coming soon”. Both repositories, the full checkpoint and a quantised NVFP4 version, were updated on Hugging Face on 29 October.

Mistral Large 4 on Terminal-Bench 4.0, public preview vs the open-weight release of 29 October 2026, as report
Mistral Large 4 on Terminal-Bench 4.0, public preview vs the open-weight release of 29 October 2026, as reported by Mistral: 28% to 41%, with 30% fewer output tokens. The 46% relative gain and the token index are our arithmetic. Source: Mistral blog post, 29 Oct 2026.

What Was Posted on X

Mistral AI on X, 29 Oct: “Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity”; it says Le Chonk ranks #1 on Cognition’s FrontierCode among open models.

Unsloth AI on X, 29 Oct: Mistral’s Le Chonk can now run locally; the 1-trillion-parameter model runs on a Mac Studio with Unsloth’s 1-bit GGUF, which it says retains 91% top-1 accuracy.

TechCrunch on X, 29 Oct: Mistral says its new model can afford organizations a kind of control over its cyber defenses that closed models can’t.

MTS on X, 29 Oct, relaying Mistral researcher Sagar Vaze: “It’s a trillion parameter model trained on 1E25 FLOPs”; “This model was trained from scratch on our own compute, which we built just outside of Paris”; “On the AA coding index, it’s exactly tied with GLM 5.3 but needs 50% less output tokens.”

Elie Bakouch posted an architecture readout of the released Mistral Large 4 weights on 29 October. In his analysis, the model uses full sparse attention, as in DeepSeek V3.2 and the GLM 5.X series, with KV cache sharing for NoPE layers after layer 35. He writes that it uses SSMax and LeRoPE, which he describes as YaRN plus a learned term per frequency, for better long context, and that there is no KV cache compression.

The Licence Carve-Out

The licence file defines “Model as a Service” as “giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints)” in a way that lets that third party control the inputs, parameters or training data. Any licensee that runs such a business “must enter into a separate commercial license with Mistral” before commercial use.

Mistral’s documentation lists its previous flagship, Mistral Large 3, under Apache 2.0, and Mistral Large 4 under “Open”. On Mistral’s own price list, Mistral Large 4 costs $1.36 per million input tokens and $4.18 per million output tokens, the same figures we reported on 28 October.

What Three Weeks of Training Changed

Mistral says “Just three weeks of additional reinforcement learning” moved the model forward on coding. Its post says that on Terminal-Bench 4.0, “Mistral Large 4 improved by 13 percentage points to 41% compared to its public preview”, “while reducing output tokens by 30%”.

By our arithmetic, 13 points on a 28% preview score is a 46% relative gain. At the unchanged $4.18 output price, 30% fewer output tokens per task means about 30% less output spend per task, on Mistral’s own figures.

On the Artificial Analysis Coding Agent Index, Mistral says the model “scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average”. The model card also says 49 billion of roughly 1 trillion parameters are active per token, about 5% by our arithmetic.

Mistral Large 4 parameters: about 1,000 billion total, 52 billion active per token including embeddings, 49 billion excluding
Mistral Large 4 has about 1 trillion parameters, of which 49 billion are active per token (52 billion including embeddings and output layers), per the Hugging Face model card. 49 of about 1,000 is about 5%: our arithmetic. Source: Hugging Face model card, 29 Oct 2026.

The Cybersecurity Pitch

Mistral says the model “scores 51% on the Artificial Analysis Cyber Index” and reports “an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models”. The post carries statements from Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik and Zscaler.

It also says that in live traffic since the preview, “we see a potential for uplift to conduct real attacks”, and that it is “modifying our guardrails accordingly”. A specialised variant for penetration testing and red-teaming goes to enterprise design partners such as Orange Cyberdefense, with access “subject to rigorous vetting and contractual safeguards”.

The Structural Read

As we read the licence, anyone can download and use the weights commercially, but a company that sells inference or post-training access to them as a service needs a separate commercial license with Mistral first.

As we read Mistral’s documentation, that is a change from its previous flagship, which it lists under Apache 2.0, to a model it lists as Open under its own licence.

By our arithmetic, at the unchanged $4.18 price per million output tokens, Mistral’s reported 30% cut in output tokens on Terminal-Bench means about 30% less output spend per task.

Mistral, 29 October 2026

“Today, we’re releasing the weights of this improved Mistral Large 4”

Three Implications

FOR COMPANIES RUNNING IT THEMSELVES The model card says the weights can be used for commercial and non-commercial use, and lists a quantised NVFP4 checkpoint that targets Blackwell GPUs.

FOR INFERENCE PROVIDERS The licence says any licensee operating a Model as a Service business must enter into a separate commercial license with Mistral; Mistral did not say what that licence costs.

FOR SECURITY TEAMS Mistral says a specialised cybersecurity variant goes only to vetted enterprise design partners, and that it is modifying its guardrails after seeing potential uplift for real attacks.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework puts it this way: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, Mistral opens the weights to every user but keeps hosted-service rights behind a separate commercial license. The framework does not discuss Mistral or Mistral Large 4.

Related coverage: Mistral Large 4 Preview: 1T Parameters, Weights Pending and Mistral Adds DeepSeek Rated Above Large 4 at Under 1/3 Price.

Update, 29 Oct: Orange Cyberdefense signs on. Orange Cyberdefense, the cybersecurity arm of French telecom group Orange, has agreed a partnership with Mistral AI to add a cybersecurity-focused version of Mistral’s latest open-weight model to its products, Reuters reported on 29 October. The rollout is “gradual and, at first, limited to select services” for corporate and public-sector clients, focused first on testing clients’ defences for vulnerabilities and investigating breaches faster, and Orange says it will run the models on computing capacity located in Europe. Financial terms were not disclosed. Orange Cyberdefense reported about EUR 1.3 billion of revenue in 2025, according to the report.

What Is Not Established

The benchmark scores are Mistral’s own reporting of its and third parties’ tests; we did not run them. The documentation page still carried a “Public Preview” label when we read it. We read only Reuters’ headline, not its article. Mistral did not say what a commercial licence for hosting costs. Ratios and percentages marked as ours are our arithmetic.

This Week In The AI Supercycle

Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls

This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.

Read the full issue →

Business Engineer Framework

The Open vs Closed Meta-Framework

A Business Engineer framework for deciding which layer of a technology stack to open and which to keep proprietary.

Read the Map of AI →

The Bottom Line

Mistral released the Mistral Large 4 weights on 29 October, as its preview promised, under a licence that lets anyone use them but requires a separate deal to sell them as a hosted service. By our arithmetic, its own figures show a 46% relative gain on Terminal-Bench in three weeks with 30% fewer output tokens.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Facts come from Mistral’s blog post of 29 October 2026, the Hugging Face model card and licence file for mistralai/Mistral-Large-4-1T-A52B, and Mistral’s documentation and price list, all read on 29 October. The 6 October preview wording and the 28 October prices are from our earlier reporting of Mistral’s own pages. We read only Reuters’ headline. We did not contact Mistral. Benchmark figures are Mistral’s. Ratios marked as ours are our arithmetic.

Sources: Mistral blog, 29 Oct 2026 · Hugging Face model card · Mistral Ecosystem Open-Weight License · Mistral documentation and pricing

Scroll to Top
FourWeekMBA