Mistral releases Large 4 model weights

Mistral made the weights of its Large 4 model available for download on October 29, which is considered one of the best open-source Chinese models in terms of code capabilities.

Large 4 requires specialized hardware

Despite being downloadable by anyone, running Large 4 requires data center equipment, which is not accessible to individuals. The model has 1,000 billion parameters and even in a compressed version, it weighs about 690 GB and needs four high-end GPUs.

Smaller models will be available soon

Mistral's vice president, Pierre Stock, announced during a press briefing on October 28 that smaller versions of the model will be available within a month, with one version expected to run on a standard laptop.

Distillation method used for smaller models

Mistral uses a process called distillation to train smaller models to mimic the behavior of the large model. This method has already been used for client companies, and the future laptop model will be from the same family as Large 4 but will not match its performance level.

Details on the smaller model remain unclear

Mistral did not specify the size of the smaller model, a firm release date beyond "within the month," or any test results. The press release only mentioned "other models of this generation in the coming months" and did not clarify what "portable" meant in the context of the model.

Laptop-compatible models already exist

Mistral already offers models that run on a laptop, such as the Ministral 3, which was released in December 2025. These models are available under the Apache 2.0 license and were created by distilling a parent model.

Mistral faces competition from Chinese models

Mistral presented the pre-release version of Large 4 on October 6, a move to re-enter the market in the face of competition from Chinese laboratories. Independent tests showed it lagging behind in code and agent tasks, prompting Mistral to promise the model weights for the end of the month.

Large 4 shows improvement in agent code

The clearest progress in the Large 4 model is in agent code, where the AI not only writes code but also autonomously chains commands in a terminal to complete a task. On the Terminal-Bench 4.0 test, Large 4 improved from 28% in the pre-release to 41%, outperforming several other models.

New reasoning modes added

Mistral added two shorter reasoning modes, "low" and "medium," to the existing "high" mode, at the request of companies that tested the model. The "high" mode now consumes 50% fewer tokens, achieved through reinforcement learning.

Cybersecurity is a key selling point

Mistral highlights cybersecurity as a key feature of Large 4. On CyberGym-E2E, a test that involves reproducing and correcting real flaws in free software, Large 4 scores 79%, matching Xiaomi's MiMo-V2.6-Pro. This performance is significantly better than models like Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.

Mistral offers restricted and unrestricted versions

During the pre-release, Mistral offered two variants of Large 4: a public version with restricted cybersecurity functions and a less restricted version reserved for security companies, selected partners, and government agencies until the model weights were published.

Mistral commits to open weights

Mistral claims that nothing guarantees Chinese laboratories will continue to publish the weights of their models, while Mistral "commits" to doing so for its own. The company hosts GLM on its own servers since September.

Closed American models still lead

Despite Mistral's efforts, closed American models like Opus 5.5 and GPT-6 Astra maintain a significant lead on Terminal-Bench, scoring between 59 and 65%. Google has further raised the bar with the release of Gemini 4 Carbon.

Large 4 is available on Hugging Face

Large 4 is available on Hugging Face, via Mistral's API, and for "certain users" of Vibe, Mistral's assistant. The model has been tested by governments and companies, though specific names were not disclosed.

Mistral has notable clients and challenges

Mistral claims more than 125 large clients, including Airbus, ASML, and HSBC. However, Ecosia announced it would abandon Mistral, citing disappointment with the model's quality and performance. Mistral now aims to accelerate development to keep up with the latest advancements.