Question 1

What is Mistral Large 4?

Accepted Answer

Mistral Large 4 is Mistral AI's largest model: a 1.05T-parameter granular Mixture-of-Experts model with 49B active parameters per token that reads text and images and writes text. It combines instant answers and step-by-step reasoning in one model.

Question 2

When was Mistral Large 4 released?

Accepted Answer

On October 6, 2026, as a public preview on the Mistral API, with same-day access on OpenRouter and Vercel AI Gateway. Full timeline: /updates.

Question 3

How many parameters does Mistral Large 4 have?

Accepted Answer

1.05 trillion in total (the launch post rounds this to 1 trillion), with 49B active per token, or 52B counting embeddings and output layers, plus a 1.6B-parameter vision encoder.

Question 4

What is the Mistral Large 4 context window?

Accepted Answer

The API serves 524,288 tokens (512K), which is also what OpenRouter, Vercel, Artificial Analysis and Vals list. Mistral's model card says 1M, so plan with 524,288.

Question 5

What is the maximum output length?

Accepted Answer

262,144 tokens per response, as listed by OpenRouter and Vercel AI Gateway. Reasoning tokens count toward output.

Question 6

Can Mistral Large 4 read images?

Accepted Answer

Yes. It accepts text and images, using a 1.6B-parameter vision encoder, and Mistral reports 63.1 on ChartQA Pro and 42.0 on Dense200 visual grounding. It does not generate images.

Question 7

Is Mistral Large 4 a reasoning model?

Accepted Answer

It is a hybrid. Set reasoning_effort to "high" to get thinking chunks before the answer, or "none" for a direct reply. In one Hacker News test, Simon Willison saw little difference between the two settings.

Question 8

Does Mistral Large 4 support function calling and JSON output?

Accepted Answer

Yes. Function calling and structured outputs work on the chat completions and conversations endpoints, and OpenRouter exposes tools, tool_choice and response_format.

Question 9

What languages does Mistral Large 4 speak?

Accepted Answer

Its training data covers more than 160 languages, including every official EU language.

Question 10

How much does the Mistral Large 4 API cost?

Accepted Answer

List price is $1.36 per 1M input tokens, $4.18 per 1M output tokens and $0.14 per 1M cached tokens. A temporary 50% sale currently makes it $0.68 / $2.09 / $0.07, and batch jobs get a further 50% off.

Question 11

Is the 50% discount permanent?

Accepted Answer

No. Mistral labels it a temporary sale price and has not announced an end date, so budget long-term usage at $1.36 / $4.18.

Question 12

Why do OpenRouter and Vercel show $0.68 / $2.09?

Accepted Answer

They pass through Mistral's sale price. OpenRouter's endpoint data carries a 0.5 discount on the $1.36 / $4.18 list price.

Question 13

Can I use Mistral Large 4 for free?

Accepted Answer

Yes. Our homepage chat gives 3 free messages a day without an account, and a free account gets 15 credits a day. One credit covers about one short reply.

Question 14

How much is the Pro plan on this site?

Accepted Answer

Pro costs $39.90 a month or $199.90 a year and includes 3,000 credits a month. The yearly plan works out to about $16.66 a month ($199.90 / 12).

Question 15

Is Mistral Large 4 open source?

Accepted Answer

Mistral announced it as an open-weight model, but the weights are not downloadable yet and the license has not been announced. Artificial Analysis currently labels it a proprietary model for that reason.

Question 16

When will the weights be on Hugging Face?

Accepted Answer

Mistral says by the end of October 2026. The repo mistralai/Mistral-Large-4.0-1T05-A52B shows an ETA of October 31, and some press reports say October 27.

Question 17

What license will Mistral Large 4 use?

Accepted Answer

Mistral has not announced it. For reference, Mistral Large 3 was released under Apache 2.0.

Question 18

Can I run Mistral Large 4 locally?

Accepted Answer

Only on a multi-GPU server, and only after the weights ship. Weights alone need about 1,050 GB at FP8 and 525 GB at 4-bit (1.05e12 parameters × bytes per parameter). Details: /run-locally.

Question 19

Does vLLM, llama.cpp or Ollama support Mistral Large 4?

Accepted Answer

Not yet. As of October 6, 2026 no Mistral Large 4 support has been announced in vLLM, SGLang, llama.cpp or Transformers.

Question 20

What is the Mistral Large 4 API model name?

Accepted Answer

mistral-large-4, with the alias mistral-large-4-0. On OpenRouter it is mistralai/mistral-large-4-0, and on Vercel AI Gateway mistral/mistral-large-4.

Question 21

Where can I use Mistral Large 4?

Accepted Answer

On the Mistral API (Mistral Studio), OpenRouter, Vercel AI Gateway and the chat on our homepage. Azure, Bedrock, Vertex and watsonx have not been announced.

Question 22

Is Mistral Large 4 available in Vibe or Le Chat?

Accepted Answer

Mistral renamed Le Chat to Vibe on August 12, 2026, and has not announced Large 4 there yet. Our homepage chat runs Mistral Large 4 today.

Question 23

How good is Mistral Large 4 at coding?

Accepted Answer

Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0. Independently, Vals AI measured 78.40% on Vibe Code Bench v1.1 (#25 of 109).

Question 24

How does it rank on independent benchmarks?

Accepted Answer

Artificial Analysis gives it 38 on Intelligence Index v4.3.2 (#64 of 225), and Vals AI gives it 48.05% on the Vals Index (#32 of 44). Its best Vals result is #6 of 75 on Harvey's Legal Agent Benchmark.

Question 25

How fast is Mistral Large 4?

Accepted Answer

Artificial Analysis measured 116.1 output tokens per second and 1.46 seconds to first token. It is also verbose: 200M output tokens to run the index against a median of 81M.

Question 26

Is Mistral Large 4 better than DeepSeek V4 Pro?

Accepted Answer

It leads DeepSeek V4 Pro 0813 on most of Mistral's agent, legal and finance charts and reads images, while independent indexes are close (38 vs 36 on Artificial Analysis). Full comparison: /vs-deepseek-v4.

Question 27

What is new compared to Mistral Large 3?

Accepted Answer

1.05T vs 675B total parameters, a reasoning mode, a 524,288-token served context instead of 256K, and 160+ training languages, at about 2.7× the list price. Full comparison: /vs-mistral-large-3.

Question 28

How does Mistral Large 4 compare with GPT-6 Astra and Claude Opus 5.5?

Accepted Answer

Both score higher on Artificial Analysis (53 and 58 vs 38) but cost far more: GPT-6 Astra is 7.4× the input and 12× the output list price, and Claude Opus 5.5 is 2.9× and 4.8×.

Question 29

What does "le Chonk" mean?

Accepted Answer

It is the nickname Mistral gave the model at launch, a nod to its size; press and social posts link it to the "Le Chaton Fat" meme.

Question 30

Where and how was Mistral Large 4 trained?

Accepted Answer

On 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Mistral has not published a knowledge cutoff date.

Fact	Mistral Large 4
Released	Public preview, October 6, 2026 (version v26.10)
API model ID	mistral-large-4 (alias mistral-large-4-0)
Parameters	1.05T total, 49B active per token, 1.6B vision encoder
Context / max output	524,288 tokens served / 262,144 tokens
Input / output	Text and images in, text out
Reasoning	reasoning_effort "high" or "none"
API price per 1M tokens	$1.36 in / $4.18 out / $0.14 cached; sale $0.68 / $2.09 / $0.07
Open weights	Promised by end of October 2026; license not announced

Topic	Figures you will see	Use this
Context window	1M in Mistral docs; 524,288 on OpenRouter, Vercel, Artificial Analysis and Vals	524,288 tokens, the limit the live API enforces
Active parameters	49B in the launch post; 52B in Mistral docs	Both: 52B counts embeddings and output layers
API price	$1.36 / $4.18 in the launch post; $0.68 / $2.09 on docs, OpenRouter and Vercel	$0.68 / $2.09 is charged now; budget long term at $1.36 / $4.18
Weights date	End of October (Mistral); Oct 31 (Hugging Face ETA); Oct 27 (press)	Between Oct 27 and Oct 31, 2026
Training GPUs	3,800 (Mistral); about 4,000 (TechCrunch, The Next Web)	3,800 Grace Blackwell GPUs
Terminal-Bench 4.0	28.3% (Mistral chart, Artificial Analysis run); 22.73% (Vals AI)	Both; the test harnesses differ

Mistral Large 4 FAQ

Mistral Large 4 at a glance

Mistral Large 4 at a glance

Where sources disagree, and which number to use

Where to read more

Frequently asked questions

More about Mistral Large 4