On October 6, 2026, a Paris laboratory dropped the biggest open-weight AI model Europe has ever built. Mistral Large 4 — the model its own creators jokingly call “Le Chonk” — packs around one trillion parameters, was trained from scratch on thousands of Nvidia’s newest chips in Mistral’s own European datacenters, and is gunning for the open-model crown currently held by Chinese labs. The preview API is live today, and the downloadable weights are expected before October ends.
If you follow AI at all, you have watched the open-weight race become a two-horse contest: closed American frontier models on one side, and increasingly dominant Chinese open models on the other. Mistral Large 4 is Europe’s bid for a third lane — and its cybersecurity scores are already turning heads. This guide breaks down what the model actually is, what its benchmarks really show, what it costs in US dollars, when you can download the weights, and who should genuinely care.
Table of Contents
- What Is Mistral Large 4?
- Why Is It Called “Le Chonk”?
- Mistral Large 4 Specs at a Glance
- Mistral Large 4 Benchmarks: What the Numbers Really Show
- The Cybersecurity Angle: Mistral Large 4’s Strongest Suit
- Mistral Large 4 Pricing: What It Costs
- Open Weights: When Can You Download Mistral Large 4?
- How Mistral Large 4 Compares to Rival Models
- How to Try Mistral Large 4 Today
- Who Should Actually Care About Mistral Large 4?
- The Honest Caveats
- Mistral Large 4 FAQ
What Is Mistral Large 4?
Mistral Large 4 (often shortened to ML4) is the new flagship AI model from Mistral AI, the Paris-based lab led by CEO Arthur Mensch. It entered public preview on Mistral’s API on October 6, 2026, and Mistral describes it as its largest and most capable model to date — a natively multimodal system built for coding, agentic workflows, and demanding enterprise work in areas like cybersecurity, finance, and law.

The headline number is the parameter count: roughly one trillion total parameters, of which only about 49 billion are “active” for any given token. That split is the whole story of how the model works. It uses a sparse mixture-of-experts design — think of it as a gigantic library where only the few shelves relevant to your question light up at a time. You get the knowledge capacity of a trillion-parameter model while paying inference costs closer to a mid-sized one. The trade-off is memory: every parameter still has to live somewhere, which is why you will not be running this on a laptop (more on that below).
It also merges two things Mistral previously kept separate: its “instruct” models (good at following directions) and its “reasoning” models (good at thinking through hard problems). The result is one hybrid model that accepts text and images as input — including dense material like engineering drawings, PDFs, and even gigapixel satellite imagery — and produces text output. Mistral says a large share of the training data covered more than 160 languages, including every official language of the European Union, which matters if you build products for a global audience.
Why Is It Called “Le Chonk”?
Here is the part of the story that made the launch genuinely fun. Back in June 2026, a joke model called “Le Chaton Fat” — the fat kitten — went viral on Reddit and X, complete with fake benchmarks and a mock claim of 30 trillion parameters. Instead of ignoring it, Mistral’s CEO Arthur Mensch played along publicly, calling it “le gros chaton” (the big kitten). When the real trillion-parameter flagship arrived in October, Mistral embraced the meme in its official launch post: “Unofficially ML4, very officially: le Chonk.” A trillion-parameter model with a cat nickname is peak 2026 AI culture — and it gives the model instant search-engine personality that drier model names lack.
Mistral Large 4 Specs at a Glance
Before the benchmarks and the pricing, here is the raw specification sheet for the model, assembled from Mistral’s announcement and model documentation. One honest note: Mistral’s own materials have shown slightly different figures at different times (1 trillion vs. 1.05 trillion total parameters; 49 vs. 52 billion active), and Artificial Analysis lists the preview’s context window lower than Mistral’s model card does. Mistral has promised full architectural disclosure when the weights ship. Where sources disagree, the table notes it.

| Specification | Mistral Large 4 |
|---|---|
| Total parameters | ~1 trillion (1.05T in model docs) |
| Active parameters per token | ~49 billion (52B including embeddings) |
| Architecture | Granular mixture-of-experts; hybrid instruct + reasoning |
| Vision encoder | 1.6 billion parameters (native multimodal) |
| Inputs / outputs | Text + images in (up to 100 images per request); text out |
| Context window | 1M tokens per Mistral’s model card; ~524K per Artificial Analysis |
| Languages | 160+, including all official EU languages |
| Training hardware | ~3,800–4,000 Nvidia Grace Blackwell GPUs, ~2 months |
| Trained / served | Mistral’s own datacenters in Europe |
| Speed | ~1.46s time-to-first-token; ~116 tokens/second output |
| Preview API pricing | $1.36 / 1M input tokens; $4.18 / 1M output tokens |
| Weights status | Promised by end of October 2026 (reports point to Oct 27–31) |
Two things stand out immediately. First, the training story: Mistral built the model from scratch in roughly two months on about 3,800–4,000 GPUs — which CEO Arthur Mensch noted is two to three times fewer chips than Chinese competitors use for comparable runs. Second, the sovereignty angle: the model was trained and is served entirely from Mistral’s own European infrastructure, which the company frames as a strategic asset for European enterprises and governments that want frontier AI under European law.

Mistral Large 4 Benchmarks: What the Numbers Really Show
Benchmarks are where AI launches get interesting — and where you need to read carefully. Nearly every headline figure for the model comes from Mistral’s own announcement, so treat them as the company’s claims until independent labs reproduce them. That said, the first independent results are in, and they paint a nuanced picture.

Mistral’s Claimed Scores
- DeepSWE v1.1 (agentic software engineering): 61.7% — a strong score on long-horizon coding tasks.
- AutomationBench (657 real business workflows): 59.9%.
- Coding Agent Index (combined agentic-coding score): 49.8%, which Mistral says puts it ahead of DeepSeek V4 Pro and Qwen3.8 Max.
- Terminal-Bench 4.0: 28.3%.
- Cybench (40 cybersecurity exercises): 93%.
- Artificial Analysis vulnerability reproduce-and-patch test: 82% — the highest score of any model on that test.
- SciCode-Verified (scientific coding): best among open models, per launch coverage.
- Dense200 (visual grounding): 42% vs. 41% for GPT-6 Astra — one of the rare cases of an open model beating a closed frontier model on any benchmark.
- Finance and legal tests: 67% on Finch/FinWorkBench, ~15% on the Harvey legal-agent benchmark, 73% on DIOR-RSVG.
What Independent Testing Adds
Artificial Analysis, the best-known independent AI benchmark outfit, published its run within hours of the launch. Their Intelligence Index scores the model at 38 — the highest score for any open model built outside China, but below seven Chinese open models, led by MiMo-V2.6-Pro at 46. On their broad aggregate index, one analysis places ML4 32nd of 44 models overall. And where Mistral claims 28.3% on Terminal-Bench 4, an independent run measured 22.73%.
The fair summary: ML4 looks like the strongest open-weight model the West has produced — genuinely competitive with the best Chinese open models on several tasks — while still trailing the closed frontier (Claude, GPT-6, Gemini) on general intelligence. Its launch materials compare it mainly against Chinese open models (DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, GLM-5.3), and that comparison set tells its own story: the open-weight frontier in 2026 is overwhelmingly a China-versus-everyone race, and Mistral is positioning itself as the one credible non-Chinese entrant.
The Cybersecurity Angle: Mistral Large 4’s Strongest Suit
If the model has a signature strength, it is cybersecurity. Mistral reports 93% on Cybench and 82% on the Artificial Analysis vulnerability reproduce-and-patch test — the highest of any model. Perhaps more striking is Mistral’s claim that Claude Opus 5.5 and GPT-6 Astra “score near zero on the same test because they refuse to perform the task” — the closed models’ safety policies block the offensive-security work that the test requires.

That strength comes with an unusual access policy. During the preview, Mistral is giving vetted cybersecurity professionals, partners, and state authorities a version of the model with reduced moderation and expanded cyber capabilities — Reuters described it as “fewer safety barriers” — so they can red-team the model in real-world settings before the weights go public. The public preview API carries more safeguards. It is a pragmatic approach to a genuinely dual-use model, and it explains why the open-weights release is staged rather than immediate.
Mistral Large 4 Pricing: What It Costs
Pricing for the preview API is refreshingly simple — and cheaper than the closed frontier, as you would expect from an open-weights-bound model:
- Input: $1.36 per million tokens
- Output: $4.18 per million tokens
- Cached input: $0.07 per million tokens
There is a wrinkle worth knowing: Mistral’s documentation has been showing a 50% launch discount — $0.68 per million input and $2.09 per million output — which Artificial Analysis describes as a two-week promotional price. If you are budgeting a project around the model, plan for the $1.36/$4.18 list rates and treat the discount as a bonus.
For context on value: at list price, the model costs roughly 5 to 11 times more per run than Anthropic’s Claude Haiku 5.5 ($0.10/$0.50 per million tokens) — but it is a far more capable model aimed at agentic coding and professional work, not quick cheap answers. Against closed frontier models charging $4–$20 per million tokens, the preview pricing is genuinely competitive. The real pricing story, though, arrives with the open weights: once you can download it, your cost becomes hardware and electricity, not tokens.
Open Weights: When Can You Download Mistral Large 4?
This is the question most developers actually care about. Mistral’s official line is that the weights will be released “by the end of the month” — end of October 2026. Reporting narrows that down: Reuters, The Next Web, and Quartz all point to October 27, while the model’s Hugging Face repository currently shows an “Expected release” of October 31, marked as an estimate. Until the files are actually published, treat the date as a target, not a promise.
Three things to know before you plan around the release:
- Open-weight is not open-source. Mistral has not yet named the license. Its earlier Mistral Large 3 shipped under Apache 2.0, but reporting suggests Large 4 will use a custom Mistral license — which could restrict commercial self-hosted deployment. Read the final terms before building a business on it.
- You cannot run it on a laptop. With ~1 trillion parameters, the model needs roughly a terabyte of GPU memory at 8-bit precision. It is a multi-GPU server model — think rented cloud clusters, not a desktop rig.
- Mistral is using the wait for red-teaming. The staged release exists so cybersecurity leaders, vetted partners, and state authorities can attack the model before everyone gets it. That is reassuring, not alarming.
How Mistral Large 4 Compares to Rival Models
Here is the model next to the models US readers are most likely choosing between right now:
- vs. DeepSeek V4 Pro and Qwen3.8 Max (Chinese open models): Mistral’s own charts show Large 4 ahead of both on its combined Coding Agent Index, and independent testing confirms it is the strongest open model built outside China. The Chinese labs still lead the open-weight pack overall on broad aggregates — but the gap is now a contest, not a coronation.
- vs. Meta’s Muse Spark Open: Meta’s recent 816-billion-parameter open release is the other Western open-weights story of the month. Both are downloadable and self-hostable; Large 4 aims higher on raw capability and agentic coding, while Spark Open’s pitch is efficiency. If you want the full open-model landscape, read our Meta Muse Spark Open explainer alongside this one.
- vs. Claude Haiku 5.5: Anthropic’s tiny model ($0.10/$0.50 per million tokens) is dramatically cheaper and faster for everyday tasks — see our Claude Haiku 5.5 guide — but it is not in the same capability class. Choose Haiku for volume; consider ML4 for hard agentic work you want to self-host.
- vs. GPT-6.1 Sol and the closed frontier: OpenAI’s and Anthropic’s top closed models still lead on general intelligence benchmarks (our GPT-6.1 Sol explainer covers OpenAI’s current efficiency play). It does not beat them head-to-head — its pitch is that it gets close enough while being downloadable, auditable, and runnable on your own hardware. For a look at the closed frontier’s current state, see our Meta Muse Wonder breakdown.
How to Try Mistral Large 4 Today
You do not have to wait for the weights. Here is how a US-based developer or curious tinkerer can put the model to work right now:
- The preview API (available now): Sign up on Mistral Studio, Mistral’s developer platform, grab an API key, and call the preview endpoint. Pay-as-you-go billing applies at the preview rates above, and the current 50% launch discount makes experimentation cheap.
- Watch the Hugging Face repository: The official weights are expected to land on Hugging Face around October 27–31. Follow the repository so you get notified the moment the checkpoint drops — early downloaders will want to verify checksums against Mistral’s official announcement.
- Plan your hardware: If you intend to self-host, start pricing multi-GPU cloud instances now. At roughly a terabyte of GPU memory at 8-bit precision, this is a serious cluster workload. Budget-conscious teams may prefer the API until demand (and quantization tooling) matures.
- Read the official announcement first: Mistral’s own launch post — “Introducing Mistral Large 4” — is the authoritative source for specs, and an independent deep-dive at DataCamp’s Mistral Large 4 guide separates the vendor claims from the independently verified results.
Who Should Actually Care About Mistral Large 4?
- Developers building coding agents: The 49.8% Coding Agent Index and 61.7% DeepSWE scores suggest it is a serious engine for agentic software work — and one you will soon be able to run yourself.
- Security teams: The cybersecurity scores are the standout. Defensive tooling, vulnerability analysis, and red-team automation are the natural first use cases.
- European enterprises and governments: A frontier-class model trained and served in Europe, under European law, is a procurement story as much as a technology story.
- Startups watching inference bills: The mixture-of-experts design means frontier-ish capability at mid-tier serving cost — attractive if your margins live or die on tokens.
- Multilingual product teams: 160+ languages with full EU coverage makes the model interesting for products that serve non-English users well.
Who can safely skip it for now? If your needs are simple Q&A, summarization, or chat at scale, cheaper models like Claude Haiku 5.5 will serve you better and cost a fraction as much. ML4 is a specialist instrument, not an everyday driver.
The Honest Caveats
No launch post tells the whole story, so here is the other side of the model:
- Most benchmarks are vendor claims. Mistral’s numbers are impressive, but independent reproduction is still thin. The gaps that exist (Terminal-Bench 28.3% claimed vs. 22.73% independently measured) are a reminder to wait for second sources.
- The docs have been inconsistent. Total parameters, active parameters, context window, and even pricing have appeared with different figures in different Mistral materials. Mistral promises full disclosure with the weights — hold them to it.
- The license is unknown. “Open weights” with a restrictive custom license is a very different proposition from Apache 2.0. Do not architect a commercial deployment until the terms are published.
- The discount will not last. The 50% launch pricing is explicitly promotional. Budget at list rates.
- It is not a laptop model. The trillion-parameter scale means serious hardware. The romance of “download and run it yourself” meets the reality of a terabyte of GPU memory.
Mistral Large 4 FAQ
What does “Le Chonk” mean?
It is Mistral’s official nickname for the model — “Unofficially ML4, very officially: le Chonk.” It traces back to a June 2026 viral joke, the fake “Le Chaton Fat” (the fat kitten) model, which CEO Arthur Mensch publicly embraced as “le gros chaton” (the big kitten).
Is Mistral Large 4 open source?
Not yet — and strictly speaking it will be open-weight, not open-source. The weights are expected by the end of October 2026 (reports point to October 27–31), but no license has been announced. Check the final license before planning commercial use.
How much does Mistral Large 4 cost?
The preview API lists $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.07. A 50% launch discount (two weeks) brings that to $0.68/$2.09 temporarily.
Can I run Mistral Large 4 on my own computer?
Realistically, no — not on a normal computer. The model needs roughly a terabyte of GPU memory at 8-bit precision, so it requires a multi-GPU server or rented cloud cluster.
Is Mistral Large 4 better than DeepSeek or Qwen?
On Mistral’s own coding-agent benchmarks, Large 4 beats DeepSeek V4 Pro and Qwen3.8 Max, and independent testing confirms it is the strongest open model built outside China. On broad aggregate indexes, several Chinese open models still rank higher overall. It is a genuine contest now, not a foregone conclusion either way.
When do the Mistral Large 4 weights release?
Mistral says “by the end of the month” (October 2026). Reuters, The Next Web, and Quartz report October 27; the Hugging Face repository shows October 31 as the current estimate.
Final Take
Mistral Large 4 is the most significant open-weight release of 2026 so far — not because it dethrones the closed frontier, but because it gives the rest of the world a third option. A trillion-parameter model, trained in Europe on European infrastructure, with best-in-class cybersecurity scores and a weights release weeks away, is a genuine event. The benchmarks deserve the usual vendor-claim skepticism, the license terms are still a blank page, and you will need a server farm to run it — but for developers, security teams, and anyone who believes AI infrastructure should not belong to two countries, Le Chonk is absolutely worth watching. We will update this guide when the weights drop.
