What Is Mistral Large 4? Benchmarks, Pricing, and Weights
Mistral released Mistral Large 4 on October 6 as a public preview, nicknamed it "le Chonk," and promised to post the weights by the end of the month.
It's the biggest model the French lab has ever built, and the launch post makes big claims: top-five cyber scores, wins over GPT-6 Astra on legal and finance work, and better coding than DeepSeek and Qwen.
I went through Mistral's announcement, model card, changelog, and pricing page, then checked each claim against Artificial Analysis, Vals AI, and LMArena on October 15. Here's what Mistral Large 4 is, how it scores outside Mistral's own charts, what it costs before and after the launch discount, and when you can download it.
TL;DR:
- What it is: Mistral AI's largest model so far, a natively multimodal mixture-of-experts model aimed at coding agents, business workflows, document and image work, and cyber defense.
- Released: October 6, 2026, as a public preview (version 26.10) in the Mistral API.
- Size: about 1 trillion total parameters (1.05 trillion on the model card), with 52 billion active per token and a 1.6-billion-parameter vision encoder.
- Context window: 1 million tokens on Mistral's own API.
- Price: $1.36 per million input tokens and $4.18 per million output tokens at list price, cut 50% to $0.68 and $2.09 during a two-week launch sale that started October 6.
- Open weights: promised by the end of October 2026. As of October 15, they aren't out, and Mistral hasn't announced the license.
- How good it is: 38 on Artificial Analysis's Intelligence Index, which would put it 8th among open-weight models once the weights ship, and well behind closed leaders like Claude Opus 5.5 (57.6). Its best independent results are in cyber defense and legal work.
What is Mistral Large 4?
Mistral Large 4 is the new flagship model from Mistral AI, released in public preview on October 6, 2026, and built to handle coding, agentic workflows, and image understanding in one model.
Mistral's launch post calls it "our largest and most capable model to date." The company trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and the preview API runs on that same hardware.
A big share of the training data was multilingual, covering more than 160 languages, including every official language of the European Union.
It's still a preview. Mistral says the reinforcement learning run behind it "is still in flight," and it expects "large and rapid improvements in the weeks and months to come," so treat every score below as a snapshot.
The specs, from Mistral's model card, changelog, and launch post:
49 billion or 52 billion active parameters?
Both numbers come from Mistral. Its Hugging Face page says Large 4 uses 49 billion active parameters per token, or 52 billion if you count the embedding and output layers.
The launch post said 49 billion for its first day and a half, then Mistral switched it to 52 billion overnight on October 7 to 8, without a note. That's why most launch coverage, and Artificial Analysis, still list 49 billion.
If "active parameters" is new to you, think of a hospital with a thousand specialists where each patient only sees a few. The whole building has to exist (that's the 1 trillion), but each visit only uses a small team (the 52 billion), which keeps answers faster and cheaper than the size suggests.
Why is it called "le Chonk"?
"Le Chonk" is a nod to a fan meme. In June, after Mistral renamed its Le Chat assistant to Vibe, fans on X and Reddit invented a fake Mistral model called "Le Chaton Fat" (roughly "the fat kitten"), complete with joke benchmarks, as Decrypt reported.
CEO Arthur Mensch played along, posting on X that "It's actually le gros chaton," French for "the big kitten." Four months later, the launch post introduced the real model as "unofficially ML4, very officially: le Chonk."
"Chonky" is informal for plump or chubby in an endearing way, usually describing an animal, per Merriam-Webster, and it can also just mean bulky or oversize. For a 1-trillion-parameter model, the name lands as a joke about size too. Mistral itself hasn't published an official origin story.
How it compares to Mistral Large 3
Mistral Large 4 is a much bigger model than Large 3. Large 3, released in December 2025, has 675 billion total parameters, 41 billion active, and a 256K context window, according to its model card.
The jump shows up in independent scores. Artificial Analysis gave Large 3 a 9 on its Intelligence Index, and Large 4 a 38.
It also costs more. Large 3 is $0.50 per million input tokens and $1.50 per million output tokens, so Large 4's list price is a little under three times higher.
Mistral Large 4 benchmarks: Mistral's numbers vs. independent tests
Mistral Large 4 looks like a frontier contender on Mistral's charts and like an upper-mid open model on independent leaderboards.
What Mistral claims
Here are the headline numbers from the launch post, with what I found when I checked each one on October 15:
Mistral also ran a blind coding review with Surge AI, where professional annotators scored outputs from 1 to 5 without knowing which model wrote them. Large 4 placed second of five at 3.74, ahead of GLM-5.3 (3.60), Kimi K3 (3.59), and GLM-5.2 (3.40), and behind only Claude Opus 5 (4.22).
What independent testing shows
On Artificial Analysis, Mistral Large 4 scores 38 on the Intelligence Index, a blend of 10 tests covering agentic work, coding, science, and long-context reasoning.
Artificial Analysis lists it as proprietary until the weights ship, but counted with open-weight models it would rank 8th, and all seven ahead of it come from Chinese labs:
- MiMo-V2.6-Pro: 46.3
- GLM-5.3: 44.8
- Kimi K3: 43.6
- GLM-5.3-Flash: 41.8
- Qwen 3.8 2.4T: 39.9
- Qwen 3.8-Flash-Next: 39.8
- DeepSeek V4.1 Flash: 39.5
The closed leaders sit well above it, including Claude Opus 5.5 at 57.6, Claude Fable 5.1 at 53.4, GPT-6 Astra at 52.7, and Gemini 4 Argon at 52.6.
Other leaderboards land in the same place. Vals AI ranks it 35th of 48 models on its overall index (48.1%), with its best result at 9th of 79 on Harvey's Legal Agent Benchmark. LMArena's crowd-voted text leaderboard had it 112th on October 13, behind DeepSeek V4 Pro.
Simon Willison, who tested it on launch day, summed it up as "back to being maybe about 6 months behind the frontier" in his first look. For a lab whose last flagship scored 9 on the same index, that's a big step.
Why the two pictures look so different
Mistral's launch charts mostly name the rivals Large 4 beats. On AA-Briefcase, Mistral names only DeepSeek V4 Pro (1,257 Elo on Artificial Analysis), while GLM-5.3, Kimi K3, MiMo-V2.6-Pro, Qwen 3.8 2.4T, and DeepSeek V4.1 Flash all score higher than Large 4's 1,393.
The coding claim works the same way. Mistral says its 49.8% Coding Agent Index beats DeepSeek V4 Pro and Qwen 3.8 Max, which is true on Artificial Analysis's coding agent table. That table also has GLM-5.3 at 53.6 and Kimi K3 at 51.9. (Large 4 isn't on that table yet, so the 49.8% is Mistral's own run of the same three tests.)
The cyber numbers hold up best. Artificial Analysis confirms the 81.7% reproduce-and-patch score, and Large 4 refused none of the tasks.
Claude Opus 5.5 and GPT-6 Astra score close to zero on that same test because they refuse 98.5% and 100% of the tasks. Those refusals come from the labs' safety policies, and they're the opening Mistral is pitching Large 4 into.
Mistral Large 4 pricing
Mistral Large 4 costs $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens at list price. Mistral's changelog says launch pricing is "50% off for 2 weeks," which brings those to $0.68, $0.07, and $2.09.
Mistral doesn't print an exact end date. Two weeks from October 6 lands around October 20, so budget anything longer-term at list price.
A few other pricing details from Mistral's pricing page and docs:
- Batch jobs: Mistral's batch API runs at half the standard price across its models.
- Regional endpoints: the EU and US regional endpoints cost 10% more than the global one.
- OpenRouter: Mistral is the only provider serving it there, at the same 50%-off sale prices.
How it compares with other open-weight models
Here's how Large 4's list price stacks up against the open-weight rivals Mistral benchmarks against, plus one closed model for scale. All prices are per million tokens from each vendor's own pricing page, checked October 15:
At list price, Large 4 lands right next to GLM-5.3 and DeepSeek V4 Pro. It costs less than half what Kimi K3 charges for input, and less than a third of Kimi's output price. Against closed models, it's cheaper still: GPT-6.1 Sol lists at $2 input and $10 output per million tokens.
DeepSeek's rates drop by half outside peak hours (the pricing page lists the windows), and DeepSeek V4.1 Flash is still the cheap option by a wide margin.
What it costs on a real job
Large 4 writes a lot, which pushes its real cost above its token price. Artificial Analysis counted 200 million output tokens to run its full index, against a median of 85 million for models in the same price tier, and most of that was reasoning.
At list prices, Artificial Analysis puts Large 4 at $1.13 per index task, compared with:
- DeepSeek V4 Pro: $0.67
- Kimi K3: $2.00
- GLM-5.3: $2.01
- GPT-6 Astra: $3.26
- Claude Opus 5.5: $5.98
A call with 100,000 input tokens and 10,000 output tokens costs about $0.18 at list price and $0.09 during the sale. If 90,000 of those input tokens come from cache, the list-price cost drops to about $0.07.
Since the API only offers two reasoning settings, "none" and "high," turning reasoning off for simple jobs is the easiest way to keep output costs down.
Mistral hasn't said whether its Vibe subscriptions include Large 4. Mistral's chat app, renamed from Le Chat to Vibe in May, costs $14.99 a month for Pro, but the launch post only points to the API preview, and Vibe's model docs still recommend Mistral Medium 3.5 with no mention of Large 4. For now, plan on paying for it per token.
What Mistral Large 4 is best at
Mistral pitches Large 4 at a handful of specific jobs, and the independent results back some of them more than others.
Cyber defense
Cyber work is where Large 4 stands out most. It ranks 5th among publicly available models on Artificial Analysis's Cyber Index, and it's far ahead of every open-weight model built outside China.
Mistral's argument is practical. Defending software often starts with proving a flaw is real, and closed models with strict safety filters tend to refuse that work.
In Mistral's internal testing, Large 4 also helped with analyzing malware, prioritizing vulnerabilities, and writing detection rules. Mistral says that on prompts asking for malicious cyber help, its refusal rate is higher than any other open model's.
Until the weights ship, Mistral is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities. Google took a similar route with Gemini 4 Argon, which went to cyber defenders before anyone else.
Business workflows and coding agents
Large 4 is built to run multi-step agent jobs that end in a finished file or a completed task. On AutomationBench, which runs 657 business workflows across apps like Gmail, Google Sheets, Slack, and Salesforce, it scores 59.9%.
For coding, it posts 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA in Mistral's tests. That's solid, but GLM-5.3 scores much higher on Terminal-Bench (41.9% vs. 26.8% in Artificial Analysis's runs), so test it against your current setup before you switch. Our roundup of AI coding agents covers the tools you'd plug it into.
Legal and finance work
Legal work is Large 4's best independent result outside cyber. It placed 9th of 79 on Harvey's Legal Agent Benchmark on Vals AI, ahead of GPT-6 Astra, Claude Opus 5.5, Kimi K3, and GLM-5.3.
It also edges out GPT-6 Astra on Vals AI's Finance Agent test, 54.7% to 53.5%, though that only puts it 24th of 78. Mistral's own demo has it working through public company filings to answer a multi-step corporate finance question.
One caveat: Mistral says Large 4 beats every open model on Harvey's benchmark, but Z.ai's GLM Legal, an open-weight model released October 12, now scores higher on Vals AI.
Documents, charts, and images
Mistral calls Large 4 a step change in how well its models understand images. Its demos show it reading engineering drawings, pulling evidence from PDFs, and scanning gigapixel satellite images for small objects.
On visual grounding, Mistral reports 42% on Dense 200, just ahead of GPT-6 Astra's 41%. On the broader MMMU-Pro test, Artificial Analysis has it at 76.4%, behind Kimi K3 (80.5%) and Claude Opus 5.5 (87.7%).
Sovereign and European deployments
For teams that need their AI under European control, Large 4 is the strongest option from a European lab. Mistral says Large 4 significantly outperforms any open-weight model developed in the US or Europe. The company will offer it in multiple regions, including a European deployment it operates end-to-end under European law, independently of other digital service providers.
Once the weights are out, it can also run on private cloud or on-premise hardware, which matters for security teams and regulated industries that need to keep data in-house.
Where it's weaker
Large 4 is less reliable on general knowledge questions. On Artificial Analysis's Omniscience test, which penalizes confident wrong answers, it scores -5.3, against 19.7 for Kimi K3 and 14.3 for GLM-5.3.
It's also wordy. Artificial Analysis puts its output volume near the top of its price tier, which means longer waits and bigger bills on hard prompts (OpenRouter measured about 70 to 80 tokens per second on October 15).
{{templates}}
Mistral Large 4 open weights and self-hosting
Mistral says the Mistral Large 4 open weights will drop by the end of October 2026, and its Hugging Face placeholder lists an expected release date of October 31. As of October 15, there are no files to download. (Some outlets reported October 27, but that date doesn't appear on any Mistral page.)
Mistral also says it will share more about the architecture, extra benchmarks, and its post-training method when the weights come out.
What license will it use?
Mistral hasn't named the license yet. Its model docs list Large 4 as "Open" and show the license as "coming soon."
Large 3 shipped under Apache 2.0, which allows commercial use and modification, but that doesn't guarantee Large 4 gets the same terms. Until the license is published, the accurate label is an open-weight model on the way, and you'll want to read the terms before building a product on it.
What hardware you'll need
Mistral Large 4 needs data center hardware, even with aggressive compression. Mistral hasn't published hardware guidance, so here's the rough math for the weights alone at 1.05 trillion parameters, before you add memory for the context window:
GPU memory totals come from NVIDIA's HGX and H200 pages. For comparison, Mistral's Large 3 model page says that 675-billion-parameter model runs in FP8 on a single node of B200s or H200s, and Large 4 is about 1.5 times bigger.
So it won't run on a laptop or a gaming PC. If you want something local today, Mistral's Ministral 3 models are the realistic option (Mistral says Ministral 3 8B fits in 12GB of VRAM in FP8). Small 4 is open too, but at 119 billion parameters it still needs serious GPU hardware.
Popular serving tools like vLLM and SGLang haven't added Large 4 support yet either. Mistral's recent open models have shipped with serving instructions for vLLM, so expect that to change when the weights land.
How to try Mistral Large 4 today
Right now, the only way to use Mistral Large 4 is through Mistral's API, either directly or through a gateway like OpenRouter. Here's the quickest path:
- Open Mistral Studio: sign up at console.mistral.ai. New accounts start in Free mode, which lets you create API keys and use a limited monthly allowance; the Limits page shows what's included.
- Test it in the playground: pick Mistral Large 4 in the Studio playground and try your own prompts before writing any code.
- Call the API: use the model ID mistral-large-4 with the chat completions or Agents endpoints. Set reasoning to "high" for hard problems and "none" for quick ones.
- Or go through OpenRouter: it's listed as mistralai/mistral-large-4-0, served only by Mistral at the same launch prices.
One gotcha: the mistral-large-latest alias still points to Large 3, so call Large 4 by its own ID. As of October 15, I couldn't find Large 4 on AWS Bedrock, Azure, or Google Vertex AI.
Should you use Mistral Large 4?
Mistral Large 4 is worth testing for security, legal, and European-hosting work, and easy to skip if you need the strongest general-purpose model.
Try Mistral Large 4 if you:
- Run security work where closed models refuse to reproduce or patch real vulnerabilities
- Need a strong model you can host yourself in Europe once the weights ship
- Handle legal or finance agent tasks and want a cheaper option than GPT-6 Astra
- Work with technical drawings, dense documents, or satellite and aerial images
Skip it for now if you:
- Want the best general model available (Claude Opus 5.5 and GPT-6 Astra score far higher)
- Need the cheapest capable open model (DeepSeek V4.1 Flash costs a fraction as much and scores slightly higher on Artificial Analysis)
- Plan to run it on a laptop or a single GPU
If you'd rather not track which model wins which benchmark each week, Lindy, an AI teammate that works in Slack, picks a model for each task, and you can choose one yourself when it matters.
{{cta}}
FAQ
Is Mistral Large 4 open source?
No, not today. Mistral Large 4 is a paid API preview, and Mistral has promised open weights by the end of October 2026, with Hugging Face showing October 31. Whether it counts as open source depends on the license, which Mistral hasn't announced.
Is Mistral Large 4 on GitHub?
Mistral Large 4 isn't on GitHub or anywhere else for download yet. Mistral's release page for the weights is a Hugging Face placeholder with an expected release date of October 31, 2026. Until then, the only way to use it is through Mistral's API.
Is Mistral Large 4 free?
Mistral Large 4 isn't free to use beyond a small allowance. The API costs $1.36 per million input tokens and $4.18 per million output tokens at list price, and half that during the launch sale. Mistral Studio's Free mode includes limited monthly usage, and the weights will be free to download once they ship, though running a 1-trillion-parameter model isn't cheap.
Which Mistral model is best?
Mistral Large 4 is Mistral's most capable model, but it's still a preview. For everyday work, Mistral's own pricing page points most tasks and coding to Mistral Medium 3.5, cost-sensitive projects to Mistral Small 4, and documents and audio to its OCR and Voxtral models.
What is the latest version of Mistral?
The latest Mistral model is Mistral Large 4, released as a public preview (version 26.10) on October 6, 2026. It's the successor to Mistral Large 3, a 675-billion-parameter model with 41 billion active parameters that came out in December 2025 and is still generally available.
How does Mistral compare to ChatGPT?
Mistral's chat app, Vibe (formerly Le Chat), costs less than ChatGPT: Pro is $14.99 a month, while ChatGPT Plus is $20 (our ChatGPT pricing guide covers the other plans). On raw ability, OpenAI's top model still leads, with GPT-6 Astra at 52.7 on Artificial Analysis against 38 for Mistral Large 4. Mistral's advantages are open weights, European hosting, and fewer refusals on security work.
What is Mistral AI used for?
Companies use Mistral AI for coding assistants, document processing, customer support, and private deployments where data has to stay in-house or in Europe. Mistral says more than 125 enterprises use its models, including Airbus, ASML, and HSBC.
Which country is Mistral AI from?
Mistral AI is from France. It was founded in Paris in 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix. In September 2026, it raised a €3 billion Series D led by Samsung, which Mistral says is the largest equity round ever raised by a European tech company.
Is Mistral Large 4 good?
Mistral Large 4 is good at specific jobs and middle of the pack overall. Independent benchmarks show a big jump from Large 3 (9 to 38 on Artificial Analysis), a top-five score among publicly available models on cyber defense, and 9th of 79 on Harvey's legal benchmark. As a general model, it would rank 8th among open-weight models on Artificial Analysis and trails the closed leaders by a wide margin.
