Mistral Large 4 Open Weights: Release Date, Specs and How to Try It Now

A French company with a fraction of OpenAI’s money just built a 1-trillion-parameter AI model, and it plans to give it away. Mistral Large 4 does security work that GPT-6 Astra refuses, will run on your own servers, and reads 160+ languages. It is not the smartest AI in the world. It might be the most important one you can own.

Last updated: 27 October 2026, 12:30 IST · Status: preview live, weights not yet released · I update this guide as soon as anything changes.

[Insert “Table of Contents by Rank Math” block here]

Mistral Large 4 in 30 seconds

Mistral Large 4, which Mistral shortens to ML4 and jokingly calls “le Chonk”, is a general-purpose AI model that can chat, write code, read documents and images, run multi-step tasks and analyse security flaws. You can rent it today through Mistral’s API. Within days, if Mistral keeps its promise, anyone will be able to download the full model and run it on their own hardware. That last part is what makes it different from ChatGPT, Claude and Gemini.

Who is Mistral AI?

Mistral AI is a Paris-based company founded in 2023 by three French researchers: Arthur Mensch, who came from Google DeepMind, and Guillaume Lample and Timothée Lacroix, who came from Meta. In three years it has gone from a small open-model lab to one of Europe’s most valuable tech companies.

When	Milestone
2023	Founded in Paris. Releases Mistral 7B and Mixtral 8x7B, small open models that punched far above their size.
Sept 2025	€1.7 billion Series C led by ASML, valuing Mistral at €11.7 billion.
2 Dec 2025	Mistral Large 3: 675 billion parameters, released the same day under the permissive Apache 2.0 licence.
2026	Le Chat becomes Vibe, a work and coding agent. New products: Studio, Forge, Compute, physics AI, OCR 4.
Sept 2026	€3 billion Series D at a valuation above €21 billion, led by Samsung Electronics.
6 Oct 2026	Mistral Large 4 public preview, announced in Abu Dhabi.
By 31 Oct 2026	Open weights promised.

Today Mistral sells a full stack, not just models: Vibe for everyday work and coding, Studio for developers, Forge for training custom models, and Compute for infrastructure. Its named industrial customers include Airbus, BMW Group and ASML, Startup Fortune reported.

Why Mistral built Large 4: the vision behind the launch

Mistral’s motto for ML4 is right in the launch headline: “Forged in Europe. Built for AI sovereignty.” The company is betting that what businesses and governments want most is not the very top benchmark score, but control.

• Control over data. With open weights, a hospital, bank or ministry can run the model inside its own building, under its own laws.
• Control over access. Mistral argues that losing access to an AI tool in the middle of a cyberattack “can itself become a critical security risk”.
• Europe can compete. ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centres, and will be served from a European region “under European law”.

“The narrative that Europe cannot compete is simply not true.”

Arthur Mensch, CEO of Mistral AI, at AI Everything Abu Dhabi (via ActuIA)

Mensch went further on the same stage: he said “99% of use cases” Mistral handles can be done with open models, and there is “no reason to pay the API margin of closed models”, ActuIA reported. In short, Mistral wants to make the expensive, locked-down AI business model of its American rivals look unnecessary.

Mistral Large 4 release date: when do the open weights drop?

Mistral’s only written commitment is this line from its official announcement: “We will release the weights by the end of the month, along with more details on the architecture, additional benchmarks, and our post-training methodology.”

• 27 October: the date Reuters and The Next Web reported at launch. Mistral never confirmed it.
• 31 October, 15:00 UTC (8:30 pm IST): the countdown on Mistral’s Hugging Face page, which carries a warning that “the release date may still change”, Forkast reported. Nearly 1,000 people had joined the waitlist by 26 October.

Until then, Mistral says it 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.” In plain words: trusted security teams are stress-testing a less-filtered version before the public gets the files.

Mistral Large 4 specs

Spec	Mistral Large 4	Mistral Large 3
Released	Preview 6 Oct 2026	2 Dec 2025
Total parameters	About 1.05 trillion	675 billion
Active per token	52 billion (Mistral)	41 billion
Vision encoder	1.6 billion parameters	—
Context window	1 million tokens (Mistral)	—
Languages	160+, including every official EU language	—
Reasoning	Hybrid: instant answers and step-by-step reasoning in one model	Instruct only at launch
Training hardware	3,800 NVIDIA Grace Blackwell GPUs	3,000 H200 GPUs
Licence	Not announced	Apache 2.0

How a 1-trillion-parameter model stays affordable

ML4 uses what Mistral’s documentation calls a “granular Mixture-of-Experts” design. Picture a hospital with hundreds of specialist doctors. You don’t see all of them for a headache; reception sends you to the two or three who fit. ML4 works the same way: out of about 1 trillion parameters, only around 52 billion “wake up” for each word. You get the knowledge of a giant model at the running cost of a much smaller one.

What can Mistral Large 4 do?

1. Cybersecurity (its standout skill)

On Artificial Analysis’s Cyber Index, which tests how well models find and fix flaws in real software, Mistral Large 4 ranks fifth of 18 models with 49.5 points. On its toughest part, CyberGym-E2E, where a model must reproduce a real vulnerability and then patch it, ML4 solved 81.7% of tasks, the best of all 18. GPT-6 Astra scored zero and Claude Opus 5.5 scored 0.8%, because they refused almost every task. Mistral also says ML4 solves 93% of Cybench, a set of 40 security-competition challenges.

2. Coding and AI agents

For coding, Mistral reports that Mistral Large 4 scores 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. In a blind test run with Surge AI, professional reviewers rated its code second of five models, behind only Claude Opus 5. On AutomationBench, 657 real business workflows across Gmail, Google Sheets, Slack and Salesforce, it scores 59.9%. CNBC still noted it “remains behind the frontier” in coding, so expect “very good”, not “best”.

3. Vision: documents, charts, drawings and satellite images

This is Mistral’s biggest jump in image understanding. Mistral Large 4 can zoom into engineering drawings to check parts, pull evidence out of PDFs, and scan giant satellite images for tiny objects, which helps disaster-response teams. On the Dense 200 visual-grounding test it narrowly beats GPT-6 Astra, 42% to 41%, by Mistral’s own numbers.

4. Office work: spreadsheets, slides, finance and law

Mistral says Mistral Large 4 can create, edit and fix complex spreadsheets and documents. In third-party tests by vals.ai on legal and financial tasks, Mistral reports it beat GPT-6 Astra, and it leads all open models on HarveyAI’s Legal Agent benchmark.

5. Science, maths and 160+ languages

Mistral calls it state of the art among open models on SciCode-Verified, a test of turning science problems into working code. A significant share of its training data spans more than 160 languages, which matters for a multilingual country like India.

6. Safety

Despite helping defenders, Mistral says ML4 refuses malicious cyber requests more often than any other open model on JailbreakBench, StrongREJECT and AgentHarm, and resists 93.3% of attacks on Lakera’s B3 prompt-injection benchmark. These are Mistral’s figures and have not been independently repeated.

Mistral Large 4 benchmarks vs GPT-6, Claude and Chinese open models

Mistral claims ML4 is “competitive with the strongest open-source models globally” and “significantly outperforming any open-weight model developed in the US or Europe”. Independent scores back the second claim firmly and the first only loosely.

Model	Maker	Download the weights?	Intelligence Index	API price (per 1M tokens, in / out)
Claude Opus 5.5	Anthropic (US)	No	58	—
GPT-6 Astra	OpenAI (US)	No	53	$10 / $50
Gemini 4 Argon	Google (US)	No	—	$2 / $10 until 30 Nov, then $4 / $20
MiMo-V2.6-Pro	Xiaomi (China)	Yes	46	—
Kimi K3	Moonshot AI (China)	Promised by 4 Nov	44	—
Mistral Large 4	Mistral (France)	Promised by 31 Oct	38	$1.36 / $4.18
Inkling Small	Thinking Machines (US)	Yes	26	—

The honest summary: Mistral Large 4 sits eighth among open-weight models, behind seven Chinese models, but 12 points ahead of the best American open model. The most impressive number is internal: Mistral Large 3 scored just 9 on the same index ten months ago. And Mistral says reinforcement learning on ML4 “is still in flight”, so today’s scores are a floor, not a ceiling. If you want to understand how OpenAI’s three GPT-6 models differ, read my GPT-6 Astra, Sol and Luna guide.

What makes Mistral Large 4 unique

1. The biggest open model outside China. Nobody in the US or Europe has published a more capable model you can download.
2. A defender that doesn’t refuse. It does real vulnerability work that GPT-6 Astra and Claude Opus 5.5 block.
3. Built and hosted in Europe. Trained in Mistral’s own data centres, with a European deployment run end-to-end under EU law.
4. One model, two modes. Quick answers and deep reasoning live in the same model instead of two separate products.
5. A foundation, not an end product. Mistral says ML4 will be the base for “a new generation of specialized and optimized Mistral models”.

What people are saying about Mistral Large 4

On Hacker News, the launch drew more than 2,000 upvotes and 1,200 comments. The conversation split into clear camps.

• Impressed: developers praised the vision and cyber scores. One wrote that the numbers were “pretty good across the board” and that people were criticising Mistral “for no reason”.
• Sceptical: others pointed out that many headline benchmarks are Mistral’s own, picked by Mistral. ActuIA noted that Mistral’s coding chart left out Kimi K3 (68%) and DeepSeek V4.1 Flash (74.2%), which both scored higher on DeepSWE.
• The “pelican test”: AI blogger Simon Willison’s famous test, asking models to draw a pelican riding a bicycle as code, sparked a debate that the test is now “saturated”, with every model drawing near-identical birds.
• Europe vs America: a large part of the thread turned into a debate over whether Europe can lead in AI at all, exactly the narrative Mistral is trying to break.

Different outlets chased different angles. CNBC and The Next Web focused on the “best open model outside China” claim. ActuIA tested that claim against independent scores. Forkast tracked the countdown and the missing licence. Startup Fortune argued the model is “almost beside the point” next to Mistral’s full business stack. One analysis also raised a “sovereignty paradox”: a model sold as European independence still runs entirely on American NVIDIA chips.

The concerns you should know about

• Safety after release. Once weights are public, anyone can retrain the model and remove its safety behaviour. ActuIA put it bluntly: the line between the public and “expanded” cyber versions “will amount to little once the weights are downloadable”. This is the same tension behind OpenAI’s recent training pause.
• Unverified benchmarks. Most coding, legal and finance numbers come from Mistral and have not been repeated by outsiders yet.
• Shifting specs. Mistral’s pages say 52 billion active parameters, while several reports cite 49 billion. Mistral lists 1 million tokens of context, but OpenRouter’s listing and Artificial Analysis’s measurement show 524K.
• The licence. Covered below, and the single most important open question.

How Mistral Large 4 helps in everyday life

Who	What they can do with it
Students	Upload a long PDF or textbook chapter and get a summary, worked examples or a quiz; photograph a graph and ask what it shows.
Creators and marketers	Turn research documents into scripts, translate captions across languages, draft slides from notes.
Small businesses	Build and fix spreadsheets, read invoices and bills from photos, draft emails and reports.
Developers	Code help, repository Q&A, terminal agents, and a model you can fine-tune and host yourself.
Banks, hospitals, government	Run frontier-level AI inside their own data centres, so sensitive data never leaves.
Security teams	Reproduce and patch vulnerabilities, analyse malware and write detection rules without being refused.

How to use Mistral Large 4

1. Developers, today: sign up on Mistral Studio, open the playground and choose Mistral Large 4. The API model ID on OpenRouter is mistralai/mistral-large-4-0.
2. Everyone else: Mistral’s app Vibe (formerly Le Chat) is where most people are likely to meet it. Mistral hasn’t confirmed which Vibe plans include ML4, so check the model picker.
3. After the weights drop: download only from Mistral’s official mistralai page on Hugging Face. Fake “leaked ML4” files are a common way to spread malware.

Can you run it on your own computer?

No, not the full model. “Only 52 billion active” does not mean it fits on a gaming PC: all trillion parameters still have to sit in memory across a rack of data-centre GPUs, and Mistral hasn’t published recommended hardware yet. For home use, Mistral’s smaller Ministral 3 models (3B, 8B and 14B) are built to run on your own device.

What Mistral Large 4 means for India

India is pushing hard for sovereign AI, with its own AI governance guidelines and a national GPU programme. An open-weight frontier model fits that goal: Indian banks, hospitals and government departments could run Mistral Large 4 inside Indian data centres instead of sending data abroad. Cheaper hosted versions from Indian cloud providers are also likely once the weights are out.

The big unknown is Indian languages. Mistral says ML4 covers 160+ languages, but nobody has published independent results for Hindi, Telugu, Tamil or Bengali yet. That is the first thing I will test once access widens.

The licence question that decides everything

“Open weights” only matters if the licence lets you use them. Mistral Large 3 shipped under Apache 2.0, which lets almost anyone use it commercially. Mistral’s newer releases use a modified MIT licence that pushes companies earning more than $20 million a month toward a paid licence or Mistral Studio. ML4’s documentation just says “Open”, with no licence named. Until that file appears, no business should plan production around it.

Mistral Large 4 price

Through Mistral’s API, Mistral Large 4 costs $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. That is about a seventh of GPT-6 Astra’s input price ($10) and a twelfth of its output price ($50). At launch, Mistral’s docs briefly showed a 50% discount ($0.68 / $2.09) with no end date.

India price check: ≈ ₹[X] / ₹[Y] per million tokens at ₹[rate] per dollar on 2026.

What happens next

• The weights and the licence should land by 31 October, with architecture details and more benchmarks.
• Specialised Mistral models built on ML4 for industries like engineering, finance and cybersecurity.
• Rapid upgrades: Mistral expects “large and rapid improvements in the weeks and months to come” as more European compute comes online.
• Competition: China’s Moonshot AI has promised Kimi K3 weights by 4 November, so the open-model race gets even tighter next week.

Watch: Mistral Large 4 explained

[Paste your YouTube or Instagram video link on an empty line below.]

FAQ

Is Mistral Large 4 free?

The weights will be free to download once released, but running a trillion-parameter model needs expensive hardware or a cloud provider. The preview API costs $1.36 per million input tokens and $4.18 per million output tokens.

Is Mistral Large 4 better than GPT-6 or Claude?

Not overall. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scores 58, GPT-6 Astra 53 and Mistral Large 4 38. ML4 wins on cybersecurity tasks, some vision tests and price, and on control, because you can run it yourself.

What is the difference between open weights and open source?

Open weights means the trained model file is published so you can run and fine-tune it. Open source usually also means the training code and data are shared. Mistral releases weights, and the licence decides how freely you can use them commercially.

Why is it called “le Chonk”?

It’s Mistral’s joke nickname for its biggest-ever model: “chonk” is internet slang for something big and chunky. Officially it is Mistral Large 4, or ML4.

Does Mistral Large 4 understand images?

Yes. It takes text and images as input and replies in text. It can read charts, scanned documents, technical drawings and satellite images.

Is it safe to download Mistral Large 4 from anywhere?

No. Only use Mistral’s official Hugging Face page or Mistral’s own platform. Unofficial “leaked” copies can contain malware.

Hari’s take

Mistral’s marketing oversells one thing: ML4 is not competing with GPT-6 Astra or Claude for the crown. But that was never the real game. What Mistral has built is the most capable AI you will be able to own outright without depending on an American or Chinese company, and for banks, governments and Indian businesses that is worth more than a few benchmark points.

My one warning: don’t celebrate the countdown, celebrate the licence. If ML4 ships under Apache 2.0 like Large 3, it becomes the default choice for anyone who wants frontier AI on their own servers. If it ships with a revenue gate, it is one strong option among many. I’ll update this guide the minute we know.

Your turn: would you trust an open model you can run yourself over ChatGPT or Gemini? Tell me in the comments.

Sources: Mistral: Introducing Mistral Large 4, Mistral docs, ActuIA, Forkast, Startup Fortune, Hugging Face community guide, Hacker News discussion, Google: Gemini 4 Argon, VentureBeat. Last updated 27 October 2026.

[mc4wp_form id="7"]