Mistral Large 4 (ML4), nicknamed “Le Chonk,” is Mistral’s largest model so far: a 1.05 trillion parameter, natively multimodal, open-weight model with 52 billion active parameters and a 1M token context window. The public preview is live on the Mistral API, and the weights are due by the end of October 2026.
What Is Mistral Large 4 (Le Chonk)?
Mistral Large 4 is an open-weight mixture-of-experts model that combines instruction following, reasoning and agentic tool use in a single model.
- Built by Mistral AI and trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters
- Accepts text and image input, with a 1.6B parameter vision encoder
- Positioned as state of the art among open models for cybersecurity, finance, law and manufacturing
- Public preview now, final weights by the end of the month
Open-weight means the trained model files can be downloaded and run on your own hardware. Mixture of experts means only about 52 billion of the 1.05 trillion parameters work on each token, which keeps running costs well below what the total size suggests.
Mistral Large 4 Specs: Parameters, Context Window and Architecture
Mistral Large 4 is a large, multimodal mixture-of-experts model built for coding, agentic workflows and cybersecurity. Here are its full specs, based on the figures on Mistral’s model docs page.
| Spec | Details |
|---|---|
| Model name | Mistral Large 4 (ML4), nicknamed Le Chonk |
| API model name | mistral-large-4 |
| Version | v26.10 |
| Total parameters | 1.05 trillion (launch post rounds to 1 trillion) |
| Active parameters | 52 billion (launch post says 49 billion) |
| Architecture | Granular mixture of experts, hybrid instruct and reasoning |
| Vision encoder | 1.6 billion parameters |
| Context window | 1M tokens |
| Input | Text and images (natively multimodal) |
| Languages | 160+, including every official EU language |
| Training hardware | 3,800 NVIDIA Grace Blackwell GPUs |
| Training location | Mistral’s own datacenters in Europe |
| Status | Public preview since October 6, 2026 |
| Weights release | By the end of October 2026 (reported as October 27) |
| License | Not announced yet |
Supported API features include structured outputs, function calling, document QnA, prefix, chat completions, batching, agents and conversations, and built-in tools.
Mistral Large 4 Release Date and Open Weights Timeline
Mistral announced ML4 and opened the public preview on October 6, 2026, with the weights due by the end of October (reported as October 27). The reinforcement learning run behind the preview is still in progress, so scores may improve before the final weights ship. Separately, cybersecurity leaders, vetted partners and state authorities are red-teaming the same model with reduced moderation and expanded cyber capabilities.
Mistral Large 4 API Pricing Per Million Tokens
The preview API is priced per million tokens, with output costing roughly three times as much as input. Cached input is the lower rate for prompt content the model has already seen in a recent request.
| Token type | Current price per 1M tokens | Struck-through list price |
|---|---|---|
| Input | $0.68 | $1.36 |
| Cached input | $0.07 | $0.14 |
| Output | $2.09 | $4.18 |
The model page crosses out the higher prices and lists the lower ones as current, which points to a preview discount. Preview pricing can change before the final release, so check the Mistral Large 4 docs page before budgeting a project.
What Real Workloads Cost
Two example workloads, calculated from the listed prices, show how the rates add up against GLM-5.3, the closest open-weight rival with public pricing on the same platform.
| Workload | ML4 (preview price) | ML4 (list price) | GLM-5.3 |
|---|---|---|---|
| 1M fresh input tokens and 200k output tokens | $1.10 | $2.20 | $2.28 |
| Agent loop: 2M fresh input, 8M cached input, 1M output | $4.01 | $8.02 | $8.32 |
Caching matters most for agents, because long tool-calling sessions resend the same context again and again. At list price ML4 is only slightly cheaper than GLM-5.3, so the real saving comes from the preview discount.
Mistral Large 4 Benchmarks: Coding, Cybersecurity and Agentic Scores
ML4’s published scores matter most in three areas: security work, coding and agents, and image understanding.
Cybersecurity
ML4 ranks among the top five models on the Artificial Analysis Cyber Index and leads open-weight models developed outside China. On the index’s reproduce-and-patch test, where a model must recreate a real vulnerability in open-source software and then fix it, ML4 scores 82%, the highest of any model tested. It also solves 93% of Cybench, a set of 40 challenges drawn from security competitions.
Several leading closed models score near zero on that reproduce-and-patch test because they refuse to attempt it. For defenders who need to prove a flaw is real before patching it, that refusal behavior is the practical case for an open-weight model. ML4 is also reported to be useful for malware analysis, vulnerability prioritisation and writing detection rules.
Coding and Agentic Work
ML4’s scores are strongest on agent and workflow tasks and more modest on pure coding.
| Benchmark | What it tests | ML4 score |
|---|---|---|
| DeepSWE v1.1 | Software engineering tasks | 61.7% |
| SWE-Atlas-QnA | Repository understanding | 59.4% |
| Terminal-Bench 4 | Complex command-line workflows | 28.3% |
| Coding Agent Index | Combined coding agent score | 49.8% |
| AutomationBench | 657 business workflows across Gmail, Google Sheets, Slack and Salesforce | 59.9% |
| AA-Briefcase | Long-horizon knowledge work | 1,393 Elo |
ML4 still trails the strongest closed models on coding, a gap Mistral has acknowledged. In a blind rating by Surge AI annotators it scored 3.74 out of 5 against 4.22 for Claude Opus 5, while beating Kimi K3 and both GLM versions (full table in the comparison section below).
Image Understanding and Safety
ML4 is strongest on visual grounding and agent security among the areas Mistral has measured.
- Visual grounding (locating objects in images): 42% on Dense 200, ahead of GPT-6 Astra at 41%
- Prompt injection (hidden instructions that hijack an AI agent): resists 93.3% of attacks on Lakera’s B3 AI Security Benchmark, with no higher score among the open models Mistral compared
How to Access and Use Mistral Large 4 (API and Open Weights)
There are three routes: the preview API today, self-hosting once the weights are published, and Mistral’s European deployment.
Method 1: Try the Preview API
- Sign in to Mistral Studio and create an API key.
- Open the playground to test the model, or call it through the API with the model name
mistral-large-4. - Send text or image input, with up to 1M tokens of context, and pay per token at the preview rates above.
Method 2: Self-Host the Open Weights
Once the weights are published, organizations can run ML4 on private cloud or on-premise infrastructure under their own policies. A 1 trillion parameter model needs serious multi-GPU hardware, so most small teams will stay on the API.
Method 3: Use Mistral’s European Deployment
Mistral says the model will be offered across multiple regions, including a European deployment it operates end to end under European law. This suits teams with data residency requirements.
Who Should Use Mistral Large 4 (and Who Shouldn’t)
ML4 is a good fit when control, security work or price matter more than having the single best coding model.
Pick ML4 if you:
- Do defensive security work and need a model that will reproduce vulnerabilities, analyse malware and write detection rules without refusing
- Need self-hosting, private cloud or European data residency
- Want a 1M token context window with image input at a lower price than GLM-5.3
- Build agents that run long, cache-heavy workflows, where the cached input rate cuts cost the most
Skip ML4 for now if you:
- Need the best possible coding quality, since the blind rating puts it behind Claude Opus 5
- Need to self-host today, because the weights are not out yet
- Need license clarity before committing, since the terms have not been announced
- Have no multi-GPU hardware and were planning to self-host a 1.05 trillion parameter model
Mistral Large 4 vs DeepSeek, Kimi and GLM: How It Compares
ML4 is positioned as competitive with the strongest open models globally, so the useful comparison is against the Chinese open-weight models most teams are weighing.
Specs and Pricing: Mistral Large 4 vs GLM-5.3
Both models offer a 1M token context window, but they differ on input type, weights status and price.
| Spec | Mistral Large 4 | Z.ai GLM-5.3 |
|---|---|---|
| Developer | Mistral AI (France) | Z.ai |
| Release | October 6, 2026 (public preview) | September 15, 2026 |
| Weights | Open weights due by end of October 2026 | Open weight |
| Input | Text and images | Text only |
| Context window | 1M tokens | 1M tokens |
| Max output | Not published | 128k tokens |
| Input price per 1M tokens | $0.68 (list $1.36) | $1.40 |
| Cached input per 1M tokens | $0.07 (list $0.14) | $0.14 |
| Output price per 1M tokens | $2.09 (list $4.18) | $4.40 |
GLM-5.3 prices are as listed on Mistral’s platform, where it is hosted without modifications.
Blind Coding Rating: Mistral Large 4 vs Kimi K3 and GLM
In a blind evaluation, Surge AI annotators rated outputs from five models on a 1 to 5 scale without seeing model names.
| Model | Score |
|---|---|
| Claude Opus 5 | 4.22 |
| Mistral Large 4 Preview | 3.74 |
| GLM-5.3 | 3.60 |
| Kimi K3 | 3.59 |
| GLM-5.2 | 3.40 |
ML4 leads the open-weight models in that rating by a modest margin, 0.14 points over GLM-5.3. On agent benchmarks it also scores ahead of DeepSeek V4 Pro 0813 on the Coding Agent Index, AutomationBench and AA-Briefcase, ahead of Qwen3.8 Max on the Coding Agent Index, and ahead of Kimi K3 and MiMo-V2.6-Pro on AutomationBench. In Mistral’s own expert review against GLM-5.3, annotators preferred ML4 in CAD and STEM and rated it level or close in finance and coding.
Frequently Asked Questions
Is Mistral Large 4 open source?
Mistral Large 4 is open-weight, which means the weights can be downloaded, but that does not guarantee a license as permissive as a traditional open source one. The weights are due by the end of October 2026, and the license terms have not been announced yet.
When will Mistral Large 4 weights be released?
Mistral says the weights will be released by the end of October 2026. News reports give the date as October 27.
How many parameters does Mistral Large 4 have?
It has 1.05 trillion total parameters and 52 billion active per token. Mistral’s launch post rounds these to 1 trillion and 49 billion.
How much does the Mistral Large 4 API cost?
The preview API currently costs $0.68 per million input tokens, $0.07 per million cached input tokens and $2.09 per million output tokens. The model page also shows $1.36, $0.14 and $4.18 as crossed-out list prices.
What is the Mistral Large 4 context window?
The context window is 1M tokens.
Does Mistral Large 4 support images?
Yes. It is natively multimodal and accepts image input, with strong results on visual grounding, documents and charts.
Does Mistral Large 4 support Telugu or Hindi?
Mistral says ML4 was trained on more than 160 languages, including every official EU language, but it has not published a full language list. Test your target language on the preview API before relying on it.
Is Mistral Large 4 free to use?
The weights will be free to download after release, subject to the license. The preview API is paid per token.
