Mistral AI has released the weights for Mistral Large 4, the model nicknamed “Le Chonk.” The October 29 release turns a model that entered public preview earlier this month into something developers and organizations can download, customize, and run on infrastructure they control.
GNW covered the earlier Mistral Large 4 preview when the weights were still promised for later in October. This release closes that gap: developers can now download the model package rather than only call a preview API.
The headline number is huge: Mistral describes Large 4 as a 1-trillion-parameter, natively multimodal mixture-of-experts model. But that number does not mean every parameter is used for every request. The company previously said the architecture activates about 52 billion parameters at a time, an approach intended to deliver large-model capability without the full inference cost of a dense trillion-parameter system.
This is not a hands-on review. The performance figures below come from Mistral and the benchmark organizations it cites, so independent testing will matter. Still, the release is significant for teams that want a capable coding or security model without sending every prompt to a third-party API.
What changed with the open-weight release
Mistral’s October 29 announcement says the company continued reinforcement learning during the three weeks after the preview. It claims the finished model improves coding, defensive cybersecurity, and reasoning efficiency while retaining multimodal input and agent-style capabilities.
The important change is access. Mistral Large 4 can now be downloaded from Hugging Face and served on private infrastructure using tools such as vLLM or SGLang. It is also available through Mistral Studio, Mistral’s API, OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints. Organizations that do not want to manage a very large model themselves therefore have managed options, while those with strict data-control requirements can keep deployment in-house.
Open weight is not quite the same as open source
Mistral calls Large 4 an open-weight model because its trained parameters are available to download. That gives developers more control than a closed API: they can inspect the model package, fine-tune it, build specialized versions, and decide where inference happens.
The license still matters. Mistral says Large 4 uses the Mistral Ecosystem Open-Weight License, a modified MIT license. It permits free use, modification, fine-tuning, and commercialization, but businesses offering the model itself as a service must obtain a separate commercial license from Mistral. Anyone planning a paid hosted service should therefore read the model card and license rather than assuming “open weight” removes every restriction.
Who can realistically run a trillion-parameter model?
The model’s mixture-of-experts design means only part of the network is active for each token, but the complete weight set is still enormous. Downloading and serving it is not a typical gaming-PC project. A practical self-hosted deployment is aimed at enterprises, research labs, cloud providers, and experienced infrastructure teams with multiple high-memory accelerators, fast interconnects, and the operational capacity to manage distributed inference.
That distinction is easy to miss. Open weights make a model controllable, not automatically cheap or simple. Smaller organizations may get the same software access but still choose a managed endpoint because the hardware, power, storage, and engineering costs of private deployment can be substantial.
Coding performance comes with vendor caveats
Mistral says Large 4 now leads open-weight models on Cognition’s FrontierCode 2.0 and Terminal-Bench 4.0. The company reports a 19.7% result on FrontierCode 2.0 and a 41% Terminal-Bench score, the latter representing a 13-point improvement over the preview.
On the Artificial Analysis Coding Agent Index, Mistral reports a 48% score, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. Those efficiency claims could matter as much as a leaderboard position because token volume affects latency and hosted inference bills.
However, benchmark setups, tool access, reasoning settings, and model-serving configurations can change results. Buyers should test Large 4 against their own repositories, languages, code-review rules, and failure cases before treating the published scores as a forecast of production performance.
Cybersecurity is the most consequential focus
The release puts unusual emphasis on defensive security work. Mistral says the model can investigate source code, reproduce vulnerabilities in isolated environments, validate flaws, and generate patches. It reports a 51% score on the Artificial Analysis Cyber Index and says the model refuses 96.6% of malicious cyber requests in its evaluation set.
That combination is valuable and inherently dual-use. A system that reasons well about vulnerabilities can assist defenders, but similar knowledge can also help attackers. Reuters reported on October 29 that Mistral improved the model after testing with governments and corporate users, while noting that downloadable weights cannot be contained like access to a hosted API.
WIRED’s same-day report also highlights the tradeoff: openness helps researchers and defenders examine behavior, but it can reduce a provider’s ability to monitor abuse. Mistral itself says it observed potential uplift for real attacks during preview traffic and is adjusting its hosted guardrails. Organizations deploying the weights will need their own access controls, logging, network isolation, human review, and authorization rules.
Three reasoning modes offer a cost-control dial
Large 4 now offers Low, Medium, and High reasoning modes. Mistral recommends Low for simple tasks, Medium for everyday work, and High—the default—for difficult jobs that need more thorough reasoning.
This gives application builders a useful control. A support workflow or document classification task may not need the same compute budget as a complex code migration or vulnerability investigation. Routing requests to the lowest adequate reasoning level can reduce response time and cost, though teams will need evaluation data to choose that threshold safely.
Hosted pricing and availability
Mistral lists hosted API pricing at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Those prices do not include the infrastructure costs of running the open weights yourself.
The model is available now through Mistral’s own services and several partners. A full technical report is still forthcoming, so some architectural and training details remain incomplete. That is another reason to separate the confirmed release details from the company’s performance claims.
Why this release matters
For most consumers, Mistral Large 4 will not be something installed on a laptop. Its impact will be indirect: it can become the engine behind coding tools, enterprise search, document analysis, private assistants, and security products.
For organizations, the appeal is more direct. Open weights allow deployment behind a company firewall, customization with proprietary data, and less dependence on a single API vendor. That can help with data residency, auditability, and long-term control. The tradeoff is that the operator—not Mistral—takes on much more responsibility for infrastructure, safety controls, updates, and misuse prevention.
Mistral Large 4 therefore lands as more than a leaderboard entry. It is a serious open-weight option for teams that have the hardware and governance to use it, while smaller developers can access the same model through hosted services. The next question is how its published results hold up in independent evaluations and real production workloads.