Mistral Large 4: Can the Model Take On Chinese Rivals?

TVS News AI, editorially supervised, presents Mistral Large 4: Can the Model Take On Chinese Rivals.

After three weeks of previewing, the French AI company Mistral AI published the weights for Large 4 on Hugging Face. The multimodal model was trained on 3,800 Nvidia accelerators and has one trillion parameters. It is therefore still too large for operation by its owner: Mistral explicitly identifies hobbyists with gaming PCs or individual Nvidia AI accelerators as a target audience that is too small.

The release ends a preview phase that involved members of an early-access customer program and specialized cybersecurity partners. These partners include ASML, Cisco, CrowdStrike, Microsoft and the European Space Agency. According to t3n, the quality of the language model improved during this phase through reinforcement learning: The model received feedback on desired results, followed by additional fine-tuning.

Large 4 Is Set to Bring Mistral Back to the Forefront in Coding and Cybersecurity

The performance claim

Large 4 is specifically aimed at Chinese open-weight models. For two benchmarks, Mistral cites specific results: In the cybersecurity benchmark CyberGym-E2E, the model ranks ahead of GLM-5.3 and Kimi K3. In the agentic coding benchmark Terminal Bench 4.0, it reaches the same level as GLM-5.3. According to t3n's assessment, Large 4 is presumably significantly better than Kimi K3 and Meta's open Muse Spark Open.

This gives the manufacturer concrete points of comparison for the first time. The caveat is important: There is not yet an independent evaluation of Large 4. Its benchmark ranking is therefore a claim by Mistral, not an overall result that has already been independently confirmed.

Why the comparison matters to Mistral

Before Large 4, Mistral marketed the medium-sized Mistral Medium 3.5 as its flagship. According to the analysis platform Artificial Analysis, however, the model lagged considerably behind Chinese competitors in some benchmarks. Mistral also opened its own computing infrastructure to competitors. According to t3n, this indicated that use of its own models had fallen short of expectations.

Large 4 is intended to address this weakness with more than just additional parameters. The focus is on tasks in which a model is expected to write code, operate tools or address security issues. The model's multimodality expands this approach to include the processing of various input formats.

Open Weights, a High Barrier to Self-Operation

The price before the release

Before the weights were released, the preview version cost US$1.4 per million input tokens and US$4.2 per million output tokens. For comparison: According to t3n, Kimi K3.1 Pro charged two dollars and ten dollars, respectively. Z.ai's API for GLM-5.3 was priced at roughly the same level as Large 4.

The prices make Large 4 attractive compared with Chinese models, but they come with a practical catch: With 3,800 Nvidia accelerators and one trillion parameters, the model is too costly to operate independently. Mistral therefore wants to release quantized versions in the coming weeks. These are also expected to include models that can be used on laptops or gaming PCs.

The role of open weights

Large 4 is available as an open-weight model. This makes the weights accessible for further development but says little about the degree of openness of the entire training and release process. t3n therefore asked Mistral at the press briefing whether the company would consider a fully transparent open-source release like Soofi or Apertus instead of open weights.

Chief Science Officer Pierre Stock sidestepped the question: “I don't think AI is such a black box as people believe. Since reasoning models have existed, it has become possible to see what the models are thinking.” Mistral relies on monitoring and decision models to increase the interpretability and transparency of larger models.

In doing so, Stock also acknowledges a limit to the presentation: Reasoning steps are visible, but verifiability could become more difficult with more complex models. Open weights and traceable decisions are therefore two separate questions.

European AI Gets a €3 Billion Boost

The financial foundation

In September 2026, Mistral raised three billion euros in fresh capital. The funds are intended to finance both the further development of its models and the expansion of its computing capacity. This timing is crucial for Large 4 because the open weights are initially affected by the high costs of training and inference.

What remains to be independently tested

The news now provides more than just a product name: 3,800 Nvidia accelerators, one trillion parameters, two benchmarks, concrete API prices and a planned series of quantized versions. But it does not yet confirm whether Large 4 works more reliably in everyday use than the leading alternatives.

The claim is nevertheless relevant for Europe's AI landscape. According to t3n, Mistral is considered the only globally competitive EU AI lab. Whether Large 4 confirms this claim through independent tests, traceable errors and realistic practical deployment remains to be seen. Until then, the score is: The benchmarks are concrete, but independent verification is missing.