Skip to content
FrontierDot

Mistral model timeline: 7B to Large 4, licenses and sizes (2023–2026)

Live timeline · last updated · 1 entry in the update log

Mistral’s model line is a license story as much as a size story. Apache 2.0, a non-production license and a research license all show up, sometimes on the same day. This page keeps the date, the grant and the parameter split that Mistral printed, and it leaves a conflict in place when two Mistral pages disagree.

As of October 21, 2026 Mistral Large 4 is in preview, announced October 6, and Mistral’s own pages give both 52 billion and about 49 billion active parameters. Weights were still described as planned, not released.

Read on 2026-10-21: Mistral’s news posts for 7B, Mixtral 8x7B, Mixtral 8x22B, Codestral, Large 2, Pixtral Large, Medium 3, Magistral, Mistral 3, Small 4, Large 4 and the NVIDIA collaboration note. Benchmark sentences are Mistral’s. We did not run the models. A Medium 3.5 URL returned 404, so it is not a node.

Timeline

  1. Mistral announces Mistral 7B. The page description calls it the best 7B model to date and names Apache 2.0. The longer article body did not come through in the fetch we kept, so this node stops at those two facts: a 7B model, Apache 2.0, on this date. Later Mistral posts still list Mistral 7B among the models that remain under Apache 2.0, which is consistent with the description, and we do not add parameter or context details the description does not give.

    Mistral: Announcing Mistral 7B

  2. Mistral announces Mixtral 8x7B under Apache 2.0, with a 32k context, and says it handles English, French, Italian, German and Spanish. The post’s speed comparison, 6 times faster inference than Llama 2 70B, is Mistral’s comparison. This is the first mixture-of-experts node in the line. The 8x7B name describes the expert layout Mistral published. We do not restate a total-parameter figure here, because the sentence we are using is the one on the announcement, and the later 8x22B post is where Mistral prints an active-versus-total split in the form we quote below.

    Mistral: Mixtral of experts

  3. Mistral releases Mixtral 8x22B, a sparse mixture-of-experts with 39 billion active parameters out of 141 billion, a 64K context, and native function calling. It is fluent in English, French, Italian, German and Spanish, and the license is Apache 2.0. An instructed version ships the same day. Mistral reports 90.8 percent GSM8K maj@8 and 44.6 percent Math maj@4 for that instructed version. The claim that it is faster than any dense 70B model is Mistral’s. Those benchmark figures are Mistral-reported. We did not rerun them.

    Mistral: Mixtral 8x22B

  4. Mistral releases Codestral, which it calls its first code model: 22 billion parameters, open weights, a 32k context, trained on 80 plus programming languages. The weights are under the Mistral AI Non-Production License, for research and testing, with commercial licenses on demand. The card is mistralai/Codestral-22B-v0.1 on Hugging Face. A new endpoint, codestral.mistral.ai, is free for a beta of 8 weeks behind a waitlist. The same model is on api.mistral.ai, billed per token. The license name matters more than the word open-weight: downloadable weights with a non-production grant are not Apache 2.0.

    Mistral: Codestral

  5. Mistral Large 2 is version 24.07, with a 128k context and 123 billion parameters. The API name is mistral-large-2407. Weights are on Hugging Face under the Mistral Research License, for research and non-commercial use. The same post lists the models that remain Apache 2.0: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Codestral Mamba and Mathstral. Codestral itself, from May, stays under the Non-Production License. Large 2 is the point where the flagship stops being Apache.

    Mistral: Mistral Large 2

  6. Mistral announces Pixtral Large, a 124 billion parameter open-weight multimodal model built on Mistral Large 2: a 123 billion parameter multimodal decoder and a 1 billion parameter vision encoder, with a 128K context. The API name is pixtral-large-latest. The same post puts Mistral Large 24.11 on Hugging Face under the Mistral Research License for research, with a commercial license from Mistral for commercial use, and calls 24.11 an upgrade over Large 24.07. Two releases share the page. The weight license for Large 24.11 is the research license, not Apache 2.0.

    Mistral: Pixtral Large

  7. Mistral announces Mistral Medium 3. The post does not describe it as open weights and does not give a parameter count, so this node has neither. The API price on the post is $0.4 input and $2 output per million tokens. Mistral says it performs at or above 90 percent of Claude Sonnet 3.7 on benchmarks, using a mix of previously reported numbers and Mistral’s harness. That comparison is Mistral’s. The post says it deploys on any cloud, including self-hosted setups of four GPUs and above, and that it is available the same day on La Plateforme and Amazon SageMaker. IBM WatsonX, NVIDIA NIM, Azure AI Foundry and Google Cloud Vertex are described as coming soon, not as available that day.

    Mistral: Mistral Medium 3

  8. Mistral announces Magistral, which it calls its first reasoning model. Magistral Small is a 24 billion parameter open-weight model under Apache 2.0, card mistralai/Magistral-Small-2506. Magistral Medium is the enterprise model, in preview in Le Chat and on La Plateforme. Mistral reports Medium at 73.6 percent on AIME 2024 and 90 percent majority@64, and Small at 70.7 percent and 83.3 percent on the same pair. Those scores are Mistral-reported. The split is the useful part: one Apache checkpoint you can download, and one hosted preview you cannot treat as the same artifact.

    Mistral: Magistral

  9. Mistral announces Mistral 3. The small models are Ministral 3 at 14 billion, 8 billion and 3 billion parameters. For each size the post lists base, instruct and reasoning variants with image understanding, under Apache 2.0. Mistral Large 3 is a sparse mixture-of-experts with 41 billion active parameters and 675 billion total, trained from scratch on 3,000 NVIDIA H200 GPUs. Base and instruct of Large 3 are under Apache 2.0. The post says the models are available that day on Mistral AI Studio, Amazon Bedrock, Azure Foundry and Hugging Face, among others. Apache on the flagship is a change from Large 2 and Large 24.11, which used the research license.

    Mistral: Mistral 3

  10. Mistral releases Mistral Small 4. The post gives 119 billion total parameters and 6 billion active per token, or 8 billion including the embedding and output layers, with 128 experts of which 4 are active, a 256k context and Apache 2.0. The API id is mistral-small-latest. Prices on the same post are $0.15 per million input tokens and $0.6 per million output tokens. Mistral says latency is down 40 percent and requests per second are 3 times Small 3. The post says Small 4 unifies Magistral, Pixtral and Devstral. A reasoning_effort control takes the values none and high. The Hugging Face collection is mistralai/mistral-small-4. The docs card is mistral-small-4-0-26-03. The 40 percent and 3 times figures are Mistral-reported.

    Mistral: Mistral Small 4

  11. Mistral puts Mistral Large 4, nicknamed Le Chonk, in public preview. The post says it is a 1 trillion parameter natively multimodal model with 52 billion active parameters. It says you can try the preview API that day on Mistral Studio, and that weights drop at the end of the month. The page’s dateModified is October 7, 2026. This node uses October 6 as the announcement day. Weights are not part of the announcement. They are a plan.

    Mistral: Mistral Large 4

  12. Mistral’s NVIDIA collaboration post describes the same Large 4 line with approximately 1 trillion total parameters and approximately 49 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs, and says weights are planned for release at the end of October. The October 6 post said 52 billion active, not approximately 49 billion. Both numbers are Mistral’s, on Mistral pages, fifteen days apart. As of this page’s last check on October 21, 2026, the weights are still described as planned, not as shipped. We do not pick one active-parameter count and hide the other.

    Mistral: NVIDIA collaboration

How we date a Mistral release

Each node uses the date on a mistral.ai news page we opened, and the link under the node is that page. If a second page adds a fact, the text says which page. We do not take a parameter count from a research index blurb when the announcement we fetched says something else, and we do not invent a count the announcement omits.

Medium 3 is the clean example of an omission. The May 7, 2025 post gives a price and a deployment floor of four GPUs. It does not give a parameter count, and it does not say the weights are open. A later URL, mistral.ai/news/mistral-medium-3-5, returned 404. The news index lists a Medium 3.5 product note dated April 29, 2026. We did not find a post body we could cite for a size, so Medium 3.5 is not a node.

The last check is October 21, 2026. Large 4’s weight drop was still a plan on that day. If Mistral publishes the weights after that, this page needs a new node rather than a silent edit to the October 6 sentence.

The license changed more than once

Apache 2.0 covers the early open models: Mistral 7B, Mixtral 8x7B and Mixtral 8x22B. Codestral in May 2024 is open weight under the Mistral AI Non-Production License, which the post limits to research and testing, with commercial terms on request. A free eight-week beta endpoint does not change that grant.

Large 2 in July 2024 moves the flagship to the Mistral Research License, research and non-commercial, and the post is explicit about what stays Apache: 7B, both Mixtral models, Codestral Mamba and Mathstral. Pixtral Large in November is open weight and multimodal. Large 24.11, announced in the same post, goes on Hugging Face under the research license again, with a separate commercial license from Mistral.

Magistral Small in June 2025 returns a reasoning checkpoint to Apache 2.0, while Magistral Medium stays a hosted preview. Mistral 3 in December 2025 puts Apache 2.0 on Ministral 3 and on the base and instruct forms of Large 3, a 675 billion parameter mixture-of-experts with 41 billion active. Small 4 in March 2026 is Apache 2.0 as well. The practical question for each row is not “is there a download.” It is which file’s license you are about to accept.

Large 4, two counts, weights still planned

On October 6, 2026 Mistral’s Large 4 post says the model is a 1 trillion parameter natively multimodal model with 52 billion active parameters, that the preview API is on Mistral Studio that day, and that weights drop at the end of the month. On October 21, 2026 the NVIDIA collaboration post says approximately 1 trillion total parameters and approximately 49 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs, with weights planned for the end of October.

Fifty-two and about forty-nine cannot both be the precise active count. They can both be what Mistral wrote. Until Mistral publishes one figure on a model card, quoting either number alone is a choice this page will not make. The availability fact is simpler. Preview API on October 6. Weights still planned on October 21. No weight node.

We also do not copy a research-page “128B” figure onto Medium 3.5 or onto Large 4. It was not on the posts above.

Prices and ids worth copying carefully

  • Medium 3, May 7, 2025: $0.4 per million input tokens and $2 per million output tokens. Not described as open weights.
  • Small 4, March 16, 2026: API id mistral-small-latest, $0.15 input and $0.6 output per million tokens, Apache 2.0.
  • Large 2 API id: mistral-large-2407. Pixtral Large API id: pixtral-large-latest.
  • Codestral: codestral.mistral.ai for the 8-week beta, and api.mistral.ai billed per token.
  • Large 4: preview API on Mistral Studio from October 6, 2026. No per-token price was on the two Large 4 pages we read.

What we left out

  • The body of the Mistral 7B announcement beyond the page description. The node is deliberately short.
  • Mistral Medium 3.5. The slug we tried returned 404, and we will not invent another one.
  • Timestral and Voxtral. They are speech and audio models. This page is the language-model line, and we are not adding them from a headline.
  • Codestral Mamba and Mathstral as their own nodes. Large 2 names them as still Apache. We did not open a separate announcement for either in this pass.
  • Any benchmark we did not see on the Mistral page cited for that node. The GSM8K, Math, AIME and latency figures above are labeled as Mistral’s.

Related timelines

The same format covers Claude, OpenAI, Gemini, Grok and Meta Muse. Labs covered in this batch are Moonshot Kimi, DeepSeek and Alibaba Qwen. The Front blog index lists every post.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

What is Mistral’s latest model?

As of October 21, 2026, Mistral Large 4 is in public preview, announced October 6, 2026. Weights were still described as planned for the end of October.

Is Mistral Large 4 open weight?

Not as of October 21, 2026. The October 6 post says weights drop at the end of the month. The October 21 post still says the release is planned.

Why do Large 4 pages disagree on size?

The October 6 post says 52 billion active parameters. The October 21 post says approximately 49 billion active parameters. Both are Mistral pages. This timeline quotes both.

Which Mistral models are Apache 2.0?

From the posts we read: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Magistral Small, Mistral 3 (Ministral 3 and Large 3 base and instruct) and Mistral Small 4. Large 2 and Large 24.11 use the Mistral Research License. Codestral uses the Non-Production License.

Update log

  • — First published. Checked Mistral news posts from Mistral 7B through the October 21, 2026 NVIDIA note.

Sources