Mistral Large 4 Pitches Open Weights With Tradeoffs for Buyers

Mistral Large 4 puts a practical question at the center of the AI model race: how much control does a business have over the system it depends on? Mistral’s answer is a planned release of downloadable weights, giving customers a potential route to operating the model beyond its hosted service.

Mistral says it opened a public preview on October 6, 2026, with weights planned for the end of October. The model, nicknamed Le Chonk, combines an ambitious cybersecurity pitch with coding results that leave stronger alternatives on the shortlist. Its appeal depends on how buyers weigh performance against control over deployment.

That distinction matters. A promising preview, a downloadable model and a practical enterprise deployment are separate milestones. Large 4 needs to be assessed against all three.

What open weights could change

The appeal of open weights is continuity. A business that can download and operate a model may be able to keep using a particular version after the developer stops offering it through an API. That could reduce the disruption of migrating an application whose behavior depends on a specific model.

Mistral frames this control as a reason to consider its technology alongside closed services. For organizations building long-lived systems, the ability to choose when to change models can matter alongside benchmark performance.

But retaining weights does not make a deployment independent of every outside dependency. The customer still needs suitable infrastructure, working software and permission to use the model under its license. Open weights can reduce one form of vendor dependence; they do not make operating the system effortless.

The preview also needs to be judged on its own terms. Mistral says it serves Large 4 on its European infrastructure. Testing that hosted service can help establish whether the model is useful, but it does not demonstrate that a buyer can run the eventual downloadable version economically.

A large model with a release still ahead

Mistral describes Large 4 as a multimodal model with roughly one trillion parameters and 49 billion active during inference. Its documentation gives the total more precisely as 1.05 trillion. The smaller active count reflects its mixture-of-experts architecture; it should not be read as the size of the complete model a customer would need to accommodate.

For comparison, Mistral Large 3 had 675 billion total parameters and 41 billion active. Large 4 therefore increases both figures, although parameter counts alone do not establish how much better a model will perform on a particular job.

The release schedule deserves similar care. Mistral’s English-language announcement targets the end of October, while its published release listings give differing dates. Treat late October as a target, with the exact date provisional. A planned release is also different from having a downloadable checkpoint ready for production evaluation.

Coding results leave room for rivals

Mistral reports 61.7% on DeepSWE v1.1, a software engineering benchmark. That is below the roughly 69% displayed for the best published configurations of GLM-5.3 and Kimi K3, and roughly 74% for selected models from OpenAI, Google and Anthropic.

The comparison needs a qualification: these figures come from published evaluations and configurations, rather than a single controlled test of every model under identical conditions. They help establish a shortlist, but do not settle which model will work best inside a particular development team.

Mistral also reports a blind coding evaluation conducted with Surge AI. Professional annotators rated outputs on a five-point scale, placing Large 4 second among five models at 3.74. Claude Opus 5 scored 4.22, while Kimi K3 scored 3.59.

Those results measure different things. A human rating of coding output should not be treated as interchangeable with a software engineering task-completion score. Together, they suggest Large 4 merits testing, while giving little reason to assume it is the strongest coding choice across the board.

Legal work produces a different comparison

On Harvey’s Legal Agent Benchmark, Vals AI lists Large 4 at a 15.83% task-pass rate, compared with 12.92% for Kimi K3, 10.83% for MiMo V2.6 Pro and 8.33% for GLM-5.3. That is a more favorable comparison for Mistral than the DeepSWE figures.

It is also a reminder to choose the benchmark that resembles the intended work. A lead on legal document tasks does not erase a gap on coding tasks. Nor does a 15.83% task-pass rate justify assuming reliable completion of an organization’s legal workload. Buyers should examine the failures as closely as the relative ranking.

Cybersecurity offers the strongest evidence

Mistral’s cybersecurity pitch has a concrete result behind it. Artificial Analysis lists Large 4 Preview at 81.7% on CyberGym-E2E-AA, leading the models displayed in that evaluation. That broadly matches the 82% figure Mistral promotes.

The benchmark asks an agent to find a vulnerability, reproduce a crash and patch the software while preserving its existing functionality tests. This is a more specific claim than saying a model is generally better at cybersecurity.

There is an important scoring limit, too. The headline result counts successful crash reproduction, a fix for that crash and passing functionality tests. Whether the patch fixes the underlying target vulnerability is tracked separately and does not determine that score. Buyers should therefore avoid reading 81.7% as a blanket vulnerability-remediation success rate.

Mistral separately reports 93% on Cybench. That remains a company-reported result for a different evaluation and should be kept separate from the independently published CyberGym-E2E-AA score.

For a security team, these results justify a closer look at defensive workflows. They do not establish that Large 4 will handle every authorized task or that another model’s failure necessarily reflects a refusal. Capability, moderation behavior and the quality of a proposed patch all need evaluation on the intended work.

Hardware and licensing shape the decision

At its stated size, Large 4 should be approached as an infrastructure project. Buyers should not plan around running the full model on an ordinary laptop or desktop. The 49 billion active parameters do not remove the need to accommodate the much larger collection of weights.

That changes the meaning of control. Self-hosting could give an organization more say over availability and model versions, but it also puts deployment capacity and operations into the purchasing decision. A smaller team may find the hosted preview easier to assess than a future self-managed installation.

Licensing is another decision point. Large 3 used Apache 2.0, but buyers should not assume Large 4 carries identical permissions. The applicable terms need to support the intended commercial use and deployment before downloadable weights become a practical procurement advantage.

Who should put Large 4 on the shortlist?

Large 4 looks most relevant to enterprises that value control over model versions and deployment, particularly those evaluating cybersecurity or specialized document workflows. Its published cyber result gives those teams a concrete reason to test it.

For teams choosing primarily on coding task performance, GLM-5.3, Kimi K3 and the stronger closed-model configurations remain necessary comparisons. Mistral’s ownership argument does not remove that performance gap in the displayed DeepSWE results.

The useful decision criteria are straightforward:

• Does Large 4 complete the organization’s actual tasks well enough to justify adoption?
• Would control over the model version materially reduce migration or availability concerns?
• Can the organization support the infrastructure and license requirements of the released weights?

The preview is enough to begin a comparison. A production commitment should depend on the released model, its terms and its performance in the buyer’s environment.