Mistral Large 4 Gives Europe a New Choice in Open-Weight AI

Mistral released the open weights of Mistral Large 4, also called Le Chonk, on October 29, letting developers download, self-host and adapt its largest model under Mistral's license after an October 6 API preview. Mistral reports Le Chonk scored 48% on the Artificial Analysis Coding Agent Index and used 47% fewer tokens per task on average than GLM-5.3, and that the improved version reached 41% on Terminal-Bench 4.0 versus 28.3% for the preview. Mistral frames the release as part of its sovereign AI strategy, positioning the model against open-weight Chinese rivals GLM-5.3 from Z.ai, Kimi K3 from Moonshot AI and DeepSeek V4.1 Flash, while Claude Opus 5.5 ranked ahead of Le Chonk in Surge Code blind evaluations.

Illustration generated with AI. On October 29, Mistral opened the weights of its largest AI model, Mistral Large 4, better known as Le Chonk. Developers could now download an improved version of the model, run it on their own infrastructure, and adapt it for their needs. For Mistral, the release was also a statement about who should control the AI a company depends on. Mistral first introduced Le Chonk through an API preview on October 6. Over the following weeks, the company continued training it, then released the improved model’s weights. According to Mistral, that work strengthened its coding and cybersecurity performance and made its reasoning more efficient. Those are the company’s claims, and we’ll look at the comparisons behind them shortly. Mistral’s announcement https://mistral.ai/news/mistral-large-4-oss/ describes the release and its results. The decision gives developers more than one way to use the model. They can call it through an API, choose a hosted deployment, or download and run it themselves under Mistral’s license. A European company building a coding assistant, for example, can consider where its code will be processed and who will operate the model before choosing a setup. Mistral presents this as part of its sovereign AI strategy: giving organizations more control over the models and infrastructure they depend on. That can matter to teams with requirements around sensitive data, providers, or how their AI systems are governed. European origins alone do not make a deployment compliant with every company’s rules. And running a model yourself brings hardware and operational costs. But Mistral’s decision adds another way for organizations to bring a capable European-developed model into their own systems. Once Le Chonk’s weights are open, a different question comes into focus: which models will teams actually compare it with? Mistral gives us a clue in its own announcement. Alongside Le Chonk, its coding comparisons feature GLM-5.3 from Z.ai, Kimi K3 from Moonshot AI, and DeepSeek V4.1 Flash. These are Chinese models competing for many of the same teams that want to deploy or adapt an open-weight model. Author-created visual based on Mistral’s reported comparisons . The token comparison is with GLM-5.3; the Surge evaluation includes Claude Opus 5.5. The numbers show why Mistral wants to be part of that conversation. The company reports that Le Chonk scored 48% on the Artificial Analysis Coding Agent Index , which evaluates agents on coding tasks, close to GLM-5.3 in its comparison. Mistral says Le Chonk used 47% fewer tokens per task on average than GLM-5.3 in that comparison. Fewer tokens can matter when agents work through many steps, though they do not tell us the final cost or whether the result is good enough for a particular project. The company also reports that the improved version reached 41% on Terminal-Bench 4.0 , which tests agents on tasks in a terminal environment. Its earlier announcement reported 28.3% for the preview. That is an improvement of roughly 13 percentage points in the reported results. The announcements do not establish that every testing condition was identical, and neither result promises the same performance in your repository. Author-created chart from Mistral’s preview and open-weight release announcements; identical test conditions are not established. One benchmark cannot predict every coding task. There is an American model in the picture, too. In the Surge Code blind evaluations cited by Mistral, human graders preferred Le Chonk’s answers over those from the compared GLM, Kimi, and DeepSeek models, while Claude Opus 5.5 ranked ahead of Le Chonk among the models evaluated. To me, these choices point to a practical contest. GLM, Kimi, and DeepSeek already offer deployable weights, so openness alone does not set Mistral apart from them. A team choosing between these models also has to weigh performance on its own tasks, license terms, provider requirements, and where it can run the system. For some European organizations, a model developed in Europe by a European company may be a meaningful part of that choice. Whether it meets their requirements depends on the actual deployment and contracts. Mistral’s emphasis on this field may also reflect a pragmatic choice about where it can compete most effectively today. In the Surge comparison the company presents, Claude Opus 5.5 is ahead of Le Chonk; the 47% token figure, meanwhile, compares Le Chonk with GLM-5.3, not Claude. That result makes the contrast worth examining, but it does not tell us that Mistral has conceded it cannot catch American labs, or establish a lasting gap across every task. What it does show is a European model entering a serious conversation with established open-weight rivals, with more than one criterion for a team to consider. The comparison is useful as a starting point, not a final ranking for every team. Agent benchmarks measure a setup: the model, the tools it uses, and the way the agent is run. Change the setup and the result can change. An AI coding model does not only need to produce a good answer. It also needs to be practical to use while someone is waiting for it. In one Hugging Face deployment example https://huggingface.co/blog/mistral-large4 , the team reported about 65 generated tokens per second for a single request on an endpoint using eight NVIDIA H200 GPUs and FP8 weights. That measurement describes a particular setup. It does not mean every deployment will reach that speed, or that a coding task will finish quickly simply because tokens appear quickly. The full task also includes the time spent reasoning, calling tools, running tests, and correcting mistakes. Speed is part of the choice, but the useful measurement is how long your complete task takes, including those steps. Suppose a security team wants an assistant to help investigate an unfamiliar vulnerability. A hosted service may impose a set of rules across many customers. A company running a model in its own environment has more room to define the surrounding access controls, review processes, and policies for its own use case. That does not mean every behavior inside the model disappears when the weights are downloaded. Mistral says Le Chonk has built-in guardrails and reports both performance on defensive security tasks and refusals on selected malicious requests. The organization controls more of the deployment around the model; changing the model’s learned behavior can require additional work and may have limits. Author-created illustration of deployment choices and their limits, informed by Mistral’s release and model card . The organization’s actual control depends on its implementation and license. There is also a practical reason for the nickname. Le Chonk has roughly a trillion parameters in total, with about 52 billion active for each token. It uses a mixture-of-experts architecture, which activates only part of the model for a given token. That reduces the work performed for each step, but the larger model still has to be made available to the system running it. Mistral’s model card https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B gives the technical details and deployment examples. Running it yourself means paying for suitable hardware, operating the service, and deciding how to manage access and updates. Hugging Face’s deployment guide https://huggingface.co/blog/mistral-large4 gives that cost of control a concrete shape. Its 2-bit variants run on an Apple Silicon Ultra computer with 512 GB of unified memory . The team reports about 15 generated tokens per second for the Q2 K variant on one M3 Ultra. Quantization compresses the model, and this is one configuration, not a universal minimum or a guarantee of unchanged quality. The same guide also describes an endpoint with eight H200 GPUs. The weights are also released under Mistral’s own open-weight license, not an unrestricted promise to use the model in any way. Mistral says businesses offering the model as a service need a separate commercial license. A team should check the actual terms for its use case. Let’s return to the company choosing a coding assistant. First, it can list its non-negotiable requirements: where its code may be processed, which providers it can use, and who must control access. Those requirements decide whether a hosted API is even an option and which deployments deserve a closer look. Then the team can take a small set of representative tasks, such as fixing an existing bug and adding a test, and run the same tasks against Le Chonk and another model it is allowed to use. Keep the agent tools and evaluation criteria as consistent as practical, and record model versions and settings. Compare the accepted result, end-to-end time, token and infrastructure costs, and the human review needed to trust the change. Repeat enough runs to see variation before drawing a conclusion. This is a proposed way to evaluate the choice, not a test I have run. If a hosted endpoint meets the team’s requirements, it may be the simpler place to start. If it does not, the self-hosted version gives the team a different option, with hardware and operational costs that belong in the comparison. I find this release interesting because Mistral does not need every team to choose Le Chonk over the strongest closed American model for the same reason. Some teams will choose a hosted model for its capabilities or convenience. Others need a strong model they can deploy and adapt under their own requirements. That is where I see Le Chonk’s opportunity. Mistral is competing with serious Chinese open-weight models while giving organizations another European-developed option. The benchmarks suggest it belongs in that conversation; they do not settle which model will work best for every team. For a company that needs more control over its AI, having another capable option to evaluate is already meaningful. The next decision is whether that control is worth the infrastructure and responsibility that come with it. Mistral Large 4 Gives Europe a New Choice in Open-Weight AI https://pub.towardsai.net/mistral-large-4-gives-europe-a-new-choice-in-open-weight-ai-574355cbf8c1 was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.