Beginner Basics

Mistral Large 4 Offers Europe a New Open-Weight AI Option

Mistral opened Large 4 weights for European teams to self-host a coding AI. Benchmarks show efficiency, but task fit, licenses, and costs need tests.

Mistral Large 4 Offers Europe a New Open-Weight AI Option

Key Takeaways

  • Mistral released open weights for Mistral Large 4 (“Le Chonk”) on October 29, giving developers and organizations a European AI model they can download, adapt, and run on their own infrastructure.
  • The model is positioned as a stronger option for coding and cybersecurity tasks, with improved reasoning efficiency and multiple deployment paths: API, hosted deployment, or self-hosted weights under Mistral’s license.
  • Benchmark comparisons show Le Chonk competing with major Chinese open-weight models such as GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, while still ranking below Claude Opus 5.5 in cited human-preference evaluations.
  • Le Chonk uses a mixture-of-experts architecture with about one trillion total parameters and roughly 52 billion active per token, enabling a wide range of deployment setups from high-end GPU servers to quantized runs on Apple Silicon machines.
  • The release adds a meaningful European open-weight option, but teams still need to evaluate performance on their own tasks, license terms, operational costs, and governance requirements rather than relying on benchmarks alone.

Meta description: Mistral opened Mistral Large 4, called Le Chonk, weights on October 29, letting developers and organizations run, adapt, and compare a European AI model.

Mistral Large 4 Offers Europe a New Open-Weight AI Option

Tags: Mistral Large 4, Le Chonk, open-weight AI, AI coding agents

Featured image: Abstract friendly illustration of a large open-weight AI model being deployed on servers and laptops, with shield and code icons, no people, logos, or text.

Mistral Large 4 open weights are now available to developers and organizations. On October 29, the French AI company released the weights of its largest model, Mistral Large 4, also known as Le Chonk, after continuing to train it over the following weeks from an API preview introduced on October 6. The company says the improved version strengthened coding and cybersecurity performance and made reasoning more efficient. For teams that need to run an AI model on their own infrastructure, the release offers a new option: download the model, adapt it, and decide where it operates.

What the release gives teams

Mistral's model can now be used through an API, through a hosted deployment, or by downloading the weights and running them under Mistral's license. A European organization building a coding assistant, for example, can choose how much control it wants over where code is processed and who runs the model. Mistral presents this as part of its sovereign AI strategy, giving organizations more control over the models and infrastructure they depend on. That matters most to teams with requirements around sensitive data, provider choices, or governance of AI systems. European development does not automatically satisfy every company rule, and self-hosting adds hardware and operations costs. Still, the release adds a European-developed model that organizations can test inside their own systems.

Where Le Chonk fits in the benchmark contest

The comparisons Mistral highlighted point to a crowded field. Alongside Le Chonk, its coding comparisons include GLM-5.3 from Z.ai, Kimi K3 from Moonshot AI, and DeepSeek V4.1 Flash, all Chinese open-weight models. Mistral reports that Le Chonk scored 48% on the Artificial Analysis Coding Agent Index, close to GLM-5.3 in the company's comparison, and used 47% fewer tokens per task on average than GLM-5.3. Lower token use can matter for agents that take many steps, but it does not reveal final cost or whether an output is acceptable for a specific project.

Mistral also says the improved model reached 41% on Terminal-Bench 4.0, which tests agents in terminal environments, up from the 28.3% reported for the preview, a rise of roughly 13 percentage points in the company's figures. The announcements do not show that every test condition was identical, and a single benchmark cannot guarantee the same result in a particular repository. Agent benchmarks measure a setup, including the model, tools, and how the agent is run; changing that setup can change the result.

Mistral Large 4 Offers Europe a New Open-Weight AI Option

The comparisons also include an American model. In the Surge Code blind evaluations cited by Mistral, human graders preferred Le Chonk's answers over those from GLM, Kimi, and DeepSeek, while Claude Opus 5.5 ranked ahead of Le Chonk among the models evaluated. The 47% token figure compares Le Chonk with GLM-5.3, not Claude. Those numbers do not mean Mistral has conceded it cannot catch American labs, and they do not establish a lasting gap across every task. They do show a European model entering a serious conversation with established open-weight rivals, with more than one criterion for a team to consider.

Because GLM, Kimi, and DeepSeek already provide deployable weights, openness alone does not separate Mistral from those rivals. Teams must weigh performance on their own tasks, license terms, provider requirements, and where the system can run. For some European organizations, a model developed in Europe may be a meaningful part of that choice, but compliance depends on the actual deployment and contracts. The focus on open-weight rivals may also reflect where Mistral can compete most effectively today.

Speed, size, and the cost of control

A coding model also needs to be usable while a person waits. In one Hugging Face deployment example, 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 number describes one setup, not every deployment, and fast token generation does not mean a full coding task finishes quickly. The full task can include reasoning, calling tools, running tests, and correcting mistakes.

The model's nickname reflects its size. 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 at each step, but the full model still needs to be available to the system running it. Mistral's model card gives technical details and deployment examples.

Hugging Face's deployment guide also shows a lower-end path: 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 configuration is not a universal minimum or a promise that quality stays unchanged. The same guide describes an endpoint with eight H200 GPUs.

Running the weights yourself means paying for suitable hardware, operating the service, and managing access and updates. The weights are released under Mistral's open-weight license, not an unrestricted promise for every use. Mistral says businesses offering the model as a service need a separate commercial license, so teams should check the actual terms for their use case. Mistral also says Le Chonk has built-in guardrails and reports performance on defensive security tasks and refusals on selected malicious requests. For example, a security team that wants an assistant to investigate an unfamiliar vulnerability may need to define access controls, review processes, and policies. Downloading the weights gives an organization more control over the deployment around the model, but changing the model's learned behavior may require additional work and can have limits.

Mistral Large 4 Offers Europe a New Open-Weight AI Option

How a team can test the choice

A company choosing a coding assistant can start by listing 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 an option and which deployments deserve closer review.

Next, the team can run a small set of representative tasks, such as fixing an existing bug and adding a test, against Le Chonk and another model it is allowed to use. Keeping agent tools and evaluation criteria as consistent as practical, recording model versions and settings, and repeating runs enough to see variation can make the comparison more useful. The team should compare the accepted result, end-to-end time, token and infrastructure costs, and the human review needed to trust the change. This is a proposed evaluation path, not a test already run.

If a hosted endpoint meets the requirements, it may be the simpler place to start. If it does not, the self-hosted version gives the team another path, with hardware and operational costs that belong in the comparison.

The release is notable because Mistral does not need every team to choose Le Chonk over the strongest closed American model for the same reason. Some teams will prefer 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 Le Chonk's opportunity lies. Mistral is competing with serious Chinese open-weight models while giving organizations another European-developed option. The benchmarks suggest it belongs in that conversation, but they do not settle which model will work best for every team.

For companies that need more control over their 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.

The original account was published in Towards AI on Medium, where readers are continuing the conversation by highlighting and responding to the story.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.