Mistral Large 4
- Provider
- Mistral AI
- Status
- Preview
- Context
- 1,000,000 tok
- Price
- $1.36 / $4.18 /MTok
Mistral Large 4 is Mistral AI’s largest model, announced on 6 October 2026 and opened the same day as a public preview on Mistral’s API (Mistral). Mistral describes it as “a 1 trillion-parameter natively multimodal model with 49 billion active parameters” and calls it ML4, or “le Chonk”. The preview is priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says it “will release the weights by the end of the month” and has not named a licence.
Artificial Analysis scored the preview 38 on its Intelligence Index v4.3.2, 64th of 225 models, at $1.13 per completed index task. That is level with Claude Sonnet 5 (38), one point below DeepSeek V4.1-Flash, GPT-5.5 and Grok 4.5 (39), and seven below GLM-5.3 and Qwen3.8-Max (45). Mistral published no SWE-bench figure.
Quick specs
| Mistral Large 4 | |
|---|---|
| Announced | 6 October 2026 |
| Status | Public preview on Mistral’s API; weights not yet released |
| Architecture | Mixture-of-experts; 1 trillion parameters, 49 billion active |
| Modalities | Text and image in, text out |
| Context window | 1M tokens (Mistral docs); Artificial Analysis lists 520K |
| Input price | $1.36 / MTok |
| Cached input | $0.14 / MTok |
| Output price | $4.18 / MTok |
| Batch | Half price |
| Languages | More than 160, per Mistral |
| AA Intelligence Index | 38 (v4.3.2) |
| Cost per AA index task | $1.13 |
| Output speed | 116 tokens a second (Artificial Analysis) |
| API model ID | mistral-large-4-0 |
| Licence | Not announced |
What Mistral Large 4 is
Mistral Large 4 is a mixture-of-experts model: it holds about 1 trillion parameters and uses about 49 billion of them for each token. Mistral’s model page gives slightly different figures from its announcement, listing 1.05 trillion total parameters, 52 billion active and a 1.6-billion-parameter vision encoder. Its predecessor, Mistral Large 3, has 675 billion parameters with 41 billion active.
The model takes text and images and returns text. Mistral calls it a hybrid instruct-and-reasoning model and says it is “natively fluent in 160+ languages”, including every official language of the European Union. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacentres in Europe.
Mistral’s documentation lists a 1M-token context window. Artificial Analysis lists 520K for the preview API. Mistral’s announcement gives no context window, maximum output or knowledge cutoff.
Benchmark performance
Independent (Artificial Analysis)
| Measure | Mistral Large 4 Preview | Context |
|---|---|---|
| Intelligence Index v4.3.2 | 38 | 64th of 225 models; Mistral Large 3 scores 9 and Mistral Medium 3.5 14 on the same index |
| Cost per index task | $1.13 | DeepSeek V4.1-Flash $0.27; Gemini 3.8 Flash $1.24 |
| Output speed | 116 tokens a second | Time to first token 1.46 seconds |
| Output tokens used in the evaluation | 200M | Against an 81M median; Artificial Analysis labels it very verbose |
Source: Artificial Analysis, independent.
Vendor (Mistral’s launch figures)
| Benchmark | Mistral Large 4 | Mistral’s comparison |
|---|---|---|
| DeepSWE v1.1 | 61.7% | None given |
| SWE-Atlas-QnA | 59.4% | None given |
| Terminal-Bench 4.0 | 28.3% | None given |
| Coding Agent Index (combined) | 49.8% | Ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max |
| AutomationBench | 59.9% | Ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro |
| AA Cyber Index, reproduce-and-patch test | 82% | “The highest of any model” |
| Cybench | 93% | None given |
| Dense 200 (visual grounding) | 42% | GPT-6 Astra 41% |
Source: Mistral, vendor-run.
Mistral says the model is “significantly outperforming any open-weight model developed in the US or Europe”. That claim compares it with open-weight models before its own weights are out. Mistral published no SWE-bench Pro or SWE-bench Verified score.
The cyber result
Mistral’s largest claim is on security work. On an Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral reports 82%, “the highest of any model”, and 93% on Cybench. Mistral adds that several closed models, “including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse to perform the task”. The comparison therefore measures willingness to do the task as well as ability.
Mistral says it is “red-teaming the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities”.
Pricing and access
The preview API costs $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14 (Mistral docs). Mistral’s pricing page says batch processing halves the price, which matches the $0.68 and $2.09 shown beside the list prices on the model page.
Access at launch is through Mistral Studio: “You can try the preview API today on Mistral Studio.” The announcement does not mention Vibe, Mistral’s assistant, or any cloud partner.
Open weights: announced, not released
Mistral’s documentation labels Mistral Large 4 an open-weight model, and the announcement says: “We will release the weights by the end of the month.” No weights were available at launch and no licence has been named. Mistral Large 3 was released under Apache 2.0, and Mistral Medium 3.5 under a modified MIT licence.
How it compares
vs Mistral Large 3
Mistral Large 3, released on 2 December 2025, has 675 billion parameters, a 256K context window, Apache 2.0 weights and a list price of $0.50/$1.50 per million tokens. Mistral Large 4 scores 38 on the Artificial Analysis index against Large 3’s 9, at nearly three times the token price.
vs GLM-5.3 and Kimi K3
GLM-5.3 scores 45 on the same index at $1.40/$4.40, close to Mistral Large 4’s price, and Kimi K3 scores 44 at $3/$15. Both have published weights. Mistral’s 61.7% on DeepSWE v1.1 is below the 66.9% Z.ai reports for GLM-5.3 and the 67.5% it lists for Kimi K3; those are separate vendor runs.
vs DeepSeek V4.1-Flash
DeepSeek V4.1-Flash scores 39, one point higher, at $0.15/$0.60 off-peak and $0.27 per completed index task, about a quarter of Mistral Large 4’s $1.13. DeepSeek V4.1-Flash’s weights are published under the MIT licence.
Known limitations
Preview, not general availability. The API is a public preview, and Mistral has given no date for general availability.
Weights and licence are pending. The open weights are promised by the end of October 2026 and no licence has been named.
No SWE-bench figure. Mistral published neither SWE-bench Pro nor SWE-bench Verified.
Verbose. Artificial Analysis measured 200M output tokens across its evaluation against an 81M median, which raises real bills above the headline rate.
Unsettled specifications. Mistral’s announcement and its model page disagree on parameter counts, and Mistral’s docs and Artificial Analysis disagree on the context window.
Frequently asked questions
When was Mistral Large 4 released?
Mistral announced Mistral Large 4 on 6 October 2026 and opened a public preview on its API through Mistral Studio the same day. The model weights were not released at launch; Mistral says they will follow by the end of October 2026.
Is Mistral Large 4 open source?
Not yet. Mistral labels it an open-weight model and says it will release the weights by the end of October 2026, but no weights were available at launch and Mistral has not named a licence.
How much does Mistral Large 4 cost?
The preview API costs $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14 and batch requests at half price. Artificial Analysis measures $1.13 per completed Intelligence Index task.
How many parameters does Mistral Large 4 have?
About 1 trillion, with 49 billion active per token, according to Mistral’s announcement. Mistral’s model page lists 1.05 trillion total, 52 billion active and a 1.6-billion-parameter vision encoder.
Is Mistral Large 4 better than Mistral Large 3?
On independent measurement, yes. Artificial Analysis scores Mistral Large 4 Preview at 38 on its Intelligence Index v4.3.2, against 9 for Mistral Large 3 and 14 for Mistral Medium 3.5. It costs $1.36/$4.18 per million tokens against Large 3’s $0.50/$1.50.
What is the context window of Mistral Large 4?
Mistral’s documentation lists 1M tokens. Artificial Analysis lists 520K for the preview API, and Mistral’s announcement does not state a figure.