News · 2026-10-07

Mistral Large 4 opens as an API preview, with weights promised by October’s end

Mistral launched Mistral Large 4 as a public hosted preview on October 6 and promised its weights by the end of October. The model gives developers a new European contender for coding, tool-using agents, and image understanding, but the downloadable files and their license are still missing. Its low advertised token prices also need to be weighed against unusually long outputs in outside evaluations.

Key facts

• Mistral describes about one trillion total parameters, with 49 billion routed per token or 52 billion including embeddings and output layers.
• The October 6 release is a hosted preview; the Hugging Face page lists October 31 as the estimated weight-release date.
• Official list pricing is $1.36 per million input tokens and $4.18 per million output tokens, before the temporary discount.
• Primary source: Mistral’s Large 4 announcement.

The distinction between something developers can query and something they can inspect is the story. Mistral says the weights will arrive “by the end of the month.” That creates a concrete release checkpoint. It does not yet give users a license to modify, redistribute, or commercially deploy a self-hosted model. The Hugging Face release page remained an upcoming release in the dossier’s checks, without downloadable weights or published license text.

Large 4 uses a mixture of experts: an enormous collection of learned components, with only some involved in processing each token. Think of a specialist organization that assigns each incoming question to a few departments. The departments doing the work are a fraction of the organization, but the organization still has to exist. Active parameters therefore describe part of the computation; they do not turn the trillion-parameter model into a 49-billion-parameter download. The existing mixture-of-experts lesson explains that distinction.

Mistral’s model documentation lists 1.05 trillion total parameters, 52 billion active, and a separate 1.6-billion-parameter vision encoder. The announcement and release page use a rounded total and explain the smaller routed count. Those descriptions can be reconciled without treating one as a different model. Current evaluator pages describe text and image input with text output. The reviewed sources do not establish audio or video support.

Context capacity is less settled. Mistral advertises one million tokens, while Artificial Analysis lists 524,000 and Vals AI lists 512,000. The dossier does not resolve whether the difference is a model ceiling versus a serving limit. Buyers should establish the limit on their actual endpoint before designing a large-document workflow around the larger number.

The two-week launch promotion cuts input and output prices to $0.68 and $2.09 per million tokens. Cached input is $0.07 during the promotion, versus a $0.14 list rate. These are temporary rates, as Mistral’s October 6 changelog makes clear.

The confirmed access route is the Mistral Studio preview interface. Official documentation lists function calling, structured output, agents, batch processing, and document question answering. The reviewed announcement does not confirm a Le Chat rollout or a named cloud-partner deployment. Weight download size cannot be established while the files are absent. A minimum or recommended graphics-memory requirement for self-hosting is unstated in the reviewed material; Mistral’s reported training infrastructure is not such a requirement.

Mistral positions Large 4 as globally competitive among open models and ahead of open-weight models developed in the United States or Europe. Those are company comparisons. More architecture and benchmark detail is promised with the weights. The reinforcement-learning run is reportedly still in progress, so the preview is also a moving evaluation target rather than a fixed, fully documented artifact.

Outside measurements complicate the launch narrative. Artificial Analysis reports 200 million output tokens across its Intelligence Index, against an 81-million-token peer median. That is a concrete warning about the cost of deliberation: a cheaper token does not help as much if the model produces many more of them. Vals reports an uneven profile, with strong relative legal-agent performance and weak results on some other tasks. Neither evaluator page establishes a separately auditable post-launch test date, so these are live published results, not verified evaluations conducted after October 6.

The large Hacker News discussion reflects both enthusiasm for a European contender and doubts about benchmark presentation, cost, and promotional pricing. These comments are reactions, not independent expert validation. The strongest counterargument to the optimistic launch story is completed-task economics: compare accuracy, output length, retries, and human review, using inference-cost principles, rather than comparing token prices alone.

The practical next step is a bounded hosted trial followed by a second assessment when the promised artifacts arrive. Mistral has shipped usable access. Whether it has also delivered a reproducible, commercially usable open-weight alternative remains a separate question with an October deadline.

Primary source, verified: read the paper →

Ground Truth. (2026, October 7). Mistral Large 4 opens as an API preview, with weights promised by October’s end. Ground Truth. https://groundtruth.day/news/mistral-large-4-preview-weights-promised-for-october.html

@misc{groundtruth:mistral-large-4-preview-weights-promised-for-october,
  title  = {Mistral Large 4 opens as an API preview, with weights promised by October’s end},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/mistral-large-4-preview-weights-promised-for-october.html}
}

Topics: models · open-weights · mistral · mixture-of-experts