News · 2026-10-30

Mistral releases Large 4 weights with a separate license for commercial model hosting

Mistral released the weights for Large 4 on October 29, giving organizations a downloadable version of its flagship model after an October 6 public preview. The release allows broad modification and internal deployment, but commercial businesses offering general access to the model must obtain a separate Mistral license. The standard checkpoint is a 1.06 TB download, making operational control a substantial infrastructure commitment.

Key facts

• Mistral announced the weights release on October 29, following further training after the preview.
• The standard FP8 model repository is listed at 1.06 TB.
• Commercial model-hosting businesses face a separate licensing condition; internal use is excluded.
• Primary sources: Mistral’s announcement, the model card, and the license.

The practical change is that customers can now keep the model’s weights under their own control. They can inspect behavior, adapt the model to private material, and choose where to operate it. Mistral’s announcement says “we’re releasing the weights,” a distinction worth preserving: this is the download milestone for an existing model family, with additional training, rather than its first appearance.

The licensing boundary is unusually consequential. The Mistral Ecosystem Open-Weight License permits use, modification, distribution, fine-tuning, and sale with attribution. It contains no general revenue cap. Its additional commercial condition applies when a licensee or affiliate operates a business providing third parties meaningful control over model inference or post-training, such as a general hosted endpoint. Products embedding the model inside specific features or harnesses are expressly excluded from that definition. The license therefore makes the intended business relationship part of the deployment decision.

An analogy is buying a machine with permission to use it in your factory and build products with it, while needing another agreement to rent out the machine itself. That is only an explanatory analogy; the actual license text controls. The distinction matters to hosting companies and to organizations deciding whether their product is a bounded feature or a general model service.

The architecture helps explain the hardware. Large 4 has roughly a trillion parameters, but routes each token through a much smaller active subset: about 49 billion parameters, or 52 billion counting embeddings and output layers. Selecting a few specialists reduces computation at each step. It does not remove the need to store the full roster of specialists. The Hugging Face file listing supplies the download figure; it should not be converted into an invented graphics-memory minimum.

For running the model, Hugging Face’s deployment guide documents standard FP8 serving on eight H200 graphics processors, each a 141 GB card. That is a published configuration, not a minimum requirement. Its endpoint recipe serves up to roughly 131,000 tokens of context in that setup, below the model’s advertised million-token capacity. The same guide describes a heavily compressed version on an Apple Silicon Ultra machine with 512 GB of unified memory, reporting about 15 generated tokens a second. Those examples carry different precision and performance tradeoffs. No single minimum memory figure applies across them.

Pricing makes hosted use another realistic option. Mistral’s model documentation lists current prices of $0.70 per million input tokens and $2.10 per million output tokens, roughly half its displayed original preview prices. Commenters report intervening price changes, but the dossier did not establish a complete official price history.

Artificial Analysis’s current comparison gives Large 4 17% on HLE-Diamond, a revised 1,000-question expert test. Its v5 announcement gives 68% for the leading model and 37% for Kimi K3.1 Pro, whose weights had not yet been released. That last qualification matters: announced open-weight status does not mean readers could already download the highest-scoring comparator. These results describe a particular no-tools exam, rather than every coding or security workflow.

Reception concentrates on what “open” should mean. The Hacker News discussion includes users who value downloadable, adaptable weights and others who say the commercial hosting condition makes the label misleading. The strongest objection concerns the ambiguity of meaningful customer control, alongside the simpler point that available weights do not imply unconditional hosting rights. Neither side establishes a community consensus.

Capability also depends on the task. Artificial Analysis’s revised evaluation shows a substantial expert-question gap, while Mistral emphasizes stronger coding and defensive-security results. Changing the test changes the comparison; it does not show a sudden loss of capability. Readers weighing the release should separate weight availability, license rights, hardware, and performance on their own work. The release materially increases deployment choice, with meaningful contractual and operational limits.

Primary source, verified: read the paper →

Ground Truth. (2026, October 30). Mistral releases Large 4 weights with a separate license for commercial model hosting. Ground Truth. https://groundtruth.day/news/mistral-large-4-weights-hosting-license.html

@misc{groundtruth:mistral-large-4-weights-hosting-license,
  title  = {Mistral releases Large 4 weights with a separate license for commercial model hosting},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/mistral-large-4-weights-hosting-license.html}
}

Topics: open-weights · mistral · licensing · inference · coding