Mistral Large 4 has arrived with a very large number attached to it: one trillion parameters. The nickname is “le Chonk”, which is silly in exactly the right way. The serious question is whether the model earns the attention. My answer is yes, with reservations large enough to deserve their own context window.
The public API preview opened through Mistral Studio on 6 October 2026. That makes Large 4 available to use, but not fully available to inspect. Mistral has promised the weights by the end of October. They were not released on launch day, so calling this an open-weight model is currently a promise about the near future, not a description of what the public can download today.
That distinction matters. Access through an API lets people try the product. Weights let people study it, run it elsewhere, test it without a vendor in the middle, and find out what its licence permits. Until the weights and licence arrive, Large 4 is a powerful preview with an unfinished public identity.
Mistral describes it as a natively multimodal model with one trillion total parameters and 49 billion active parameters. Its documentation gives a different figure, around 52 billion, because the two materials appear to use different counting conventions. That is not necessarily a contradiction in the engineering, but it is a contradiction in the headline numbers. Readers should not treat every parameter figure as directly comparable.
The context window is also unsettled. Artificial Analysis reports roughly 524,000 tokens. Mistral’s own material implies up to one million, while another source gives 512,000. Those figures are close enough to suggest a very large working memory and different enough that I would not build a product around the million-token claim without checking the exact endpoint and limits.
The pricing is more immediately practical. The preview is shown at about $1.36 per million input tokens and $4.18 per million output tokens, with a 90 percent discount for cached input. One source lists a lower promotional rate of $0.68 per million input tokens and $2.09 per million output tokens. That cheaper figure should be treated as temporary. The ordinary preview pricing is already competitive for a model aimed at long documents, code, multimodal work, and agentic workflows, but verbose output can quietly turn a cheap request into an expensive one.
And Large 4 is very verbose. Artificial Analysis measured it as fast, but ranked it 105th of 225 models for verbosity. During its Intelligence Index testing, it consumed 200 million output tokens. That is not a minor personality quirk. It affects latency, cost, readability, and every system that has to process the answer afterwards. A model that reaches a good result quickly but insists on narrating the entire journey is still making somebody pay for the tour.
The independent scores are mixed rather than triumphant. Artificial Analysis gives Large 4 an Intelligence Index of 38, ranking it 64th out of 225 models. Its Cyber Index score is 50, and the reported cost per Intelligence Index task is $1.13. Those results suggest a capable and useful model, especially where speed and breadth matter. They do not establish universal dominance.
Mistral’s own claims are more ambitious. The company says Large 4 outperforms any open-weight model developed in the United States or Europe, and competes with the strongest open-source models globally. It also points to coding, agentic workflows, cybersecurity, finance, and multimodal understanding. Those are vendor claims, not independent conclusions. Several headline coding benchmarks were evaluated privately by Artificial Analysis before public release, which means they are not yet reproducible on a public leaderboard. Enthusiasm is reasonable. Certainty is not.
The same caution applies to Mistral’s interpretation of a reproduce-and-patch test. Mistral reportedly argues that other frontier models score near zero because they refuse the task. That may explain the result, but it has not been independently confirmed. A benchmark can measure useful behavior while still requiring careful inspection of what was asked, what counted as refusal, and whether the comparison was fair.
The biggest unanswered question is the licence. On launch day, there were no weights, no licence text, and no model card on Hugging Face. Mistral’s EU AI Act documentation reportedly refers to proprietary and bespoke terms. That is not enough information for developers who need to know whether they can fine-tune, redistribute, deploy commercially, or build on the model without negotiating a separate set of conditions.
This is why I see Large 4 as an important release, but not yet a settled verdict on Mistral’s open model strategy. The architecture sounds ambitious. The API is available. The price may be attractive. The independent testing shows real capability, and the speed could matter more in daily work than a glamorous leaderboard position.
But the release is asking the public to evaluate a model before the public can inspect the model. The weights are pending. The licence is unnamed. The context limit is disputed. Some benchmark evidence is private. Those are not footnotes. They are the terms of the argument.
When the weights arrive, the conversation can become more concrete. People can test the claims, compare the counting conventions, examine the licence, and see whether Large 4 is genuinely useful outside Mistral’s preferred demonstrations. Until then, “le Chonk” is a promising preview, not a completed open release.
That may sound ungenerous. It is not. Mistral has put forward something worth taking seriously. The fair response is to take the whole thing seriously, including the parts that are still missing.
Sources
Mistral’s announcement of Mistral Large 4
The Next Web’s coverage of the launch
SiliconANGLE’s report on the launch and roadmap