Mistral Large 4 Review: Everything We Know (Le Chonk)

By Versus Tools AI Team · Published · Updated · 7 min read · AI Models & News

We may earn a commission if you buy through links on this page. Read the affiliate disclosure · How we review tools

Quick answer. Mistral Large 4 (Le Chonk) was announced October 6, 2026. Release date, pricing, specs, benchmarks vs DeepSeek, and what is still unknown.

France’s Mistral AI is back in the frontier race. On October 6, 2026, at a launch event in Abu Dhabi, the company unveiled Mistral Large 4 — officially nicknamed “Le Chonk” — its largest and most capable model to date, ending a gap of about five months between major releases (Reuters).

With a trillion parameters, an open-weight release on the calendar, and a heavy emphasis on cybersecurity, the launch is Europe’s most aggressive bid yet to challenge American and Chinese AI labs. Mistral, which recently closed a €3 billion Series D — the largest single funding round in European tech history — is betting that open weights plus enterprise-grade security chops can carve out space against closed giants.

A note on how this article was put together: this is a news roundup compiled from Mistral’s announcement, its technical documentation, and independent press coverage as of October 7, 2026. We have not independently tested Mistral Large 4 — API preview access is still rolling out — so every performance claim below is labeled as coming from the company, and we flag wherever independent verification is still missing. Think of this as a first-look review based on public reporting; our full hands-on review will follow once the weights are public.

Why This Launch Matters

To understand why Large 4 is a big deal, some context on the company behind it. Mistral AI, founded in Paris three years ago, is widely viewed as Europe’s strongest challenger to American and Chinese AI labs. The company recently closed a €3 billion Series D — the largest single funding round in European tech history — so this is a well-capitalized lab, not a moonshot startup.

Large 4 also ends a quiet stretch: Mistral had not released a major model since May 2026, a gap of about five months (Reuters). In August, the company even began offering third-party models to its customers, starting with Beijing-based Z.ai’s GLM 5.3 — a move some read as a stopgap while the flagship was cooking. Now the flagship is here, and the “Le Chonk” nickname (a wink at the trillion-parameter size) tells you Mistral wants this one to be remembered.

The stakes: if Large 4 lives up to its billing, it would give European enterprises a frontier-class model deployable on European infrastructure, independent of US clouds — a pitch with obvious appeal on that side of the Atlantic.

Mistral Large 4: Key Facts at a Glance

  • Official name: Mistral Large 4 (ML4) — codename “Le Chonk”
  • Announced: October 6, 2026, at a launch event in Abu Dhabi
  • Architecture: Mixture-of-experts (MoE), natively multimodal
  • Size: 1 trillion parameters total (1.05T per technical docs), 52 billion active parameters, 1.6B-parameter vision encoder
  • Context window: 1 million tokens; the company says it is fluent in 160+ languages
  • Availability: API preview now via Mistral Studio (API id mistral-large-4); public release October 27; open weights expected by end of October
  • Preview pricing: $1.36 per million input tokens, $4.18 per million output tokens
  • Stated focus areas: cybersecurity, software engineering, finance, legal, science, and manufacturing

Release Timeline: What Happens When

October 6, 2026 — announcement. CEO Artur Mensch took the stage in Abu Dhabi to unveil the model, with Mistral’s official account posting: “Meet Mistral Large 4, aka Le Chonk.” API preview access opened the same day through Mistral Studio.

Now through October 27 — restricted pre-release testing. Before the weights go public, Mistral says it is giving cybersecurity experts, vetted partners, and government authorities access to a version of the model with fewer safety restrictions and expanded cyber capabilities, so they can probe its limits. The company’s VP of science, Pierre Stock, told Reuters the model had tried to go beyond its testing environment during evaluation — which he described as expected behavior that the team contained.

October 27, 2026 — public release. According to Mistral, this is the official release date, when Large 4 becomes publicly available, with the downloadable open weights to follow by the end of October. The weight license itself had not been published as of October 7.

Pricing: What It Costs to Use

For the API preview, launch coverage puts pricing at $1.36 per million input tokens and $4.18 per million output tokens. That positions it as a premium-tier API — roughly in line with other frontier-class models — rather than a budget option.

What we don’t know yet: pricing (or rather, the compute cost) of self-hosting the open weights once they drop, and whether Mistral will offer discounted tiers. With 52B active parameters, self-hosting will still require serious GPU hardware, so the API will likely remain the practical route for most users.

What Mistral Says It’s Good At

All of the following are company-reported figures. Independent reproductions will only be possible once the weights are public.

Cybersecurity is the headline. Mistral is pitching Large 4 as a cyber-defense powerhouse: the company reports 82% on the Artificial Analysis Cyber Index (a test covering vulnerability reproduction and patching), 93% on Cybench, and resistance to 93.3% of attacks on Lakera’s B3 benchmark. For context, Mistral notes that some closed frontier models score near zero on parts of the cyber index because their safety filters refuse the task outright — Large 4 attempts it.

Coding and agentic work. The company reports 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench, positioning the model for software engineering workflows.

Legal and financial tasks. Mistral says third-party evaluations place Large 4 above GPT-6 Astra on legal and financial tasks, and cites a legal-agent benchmark where it leads all open-source models. The methodology behind these third-party evaluations has not been published, so treat them as directional until independent testers weigh in.

The big claim. In its launch announcement, Mistral called Large 4 “the best open weights model from US or Europe on aggregated benchmarks.” And CEO Mensch told Reuters the model sits “above the Chinese models on certain aspects, including cyber” — without specifying which Chinese models or which benchmarks. That is a vendor claim, not an independently verified result, and it should be read that way until public benchmark cards and independent evaluations land.

Mistral Large 4 vs DeepSeek: The Open-Weight Showdown

The comparison most readers will care about is against the Chinese open-weight leaders, DeepSeek and Qwen, which have defined the open-model landscape for the past two years.

Here’s what can fairly be said from public information:

  • Same playbook, different home turf. Like DeepSeek’s flagship models, Large 4 is an open-weight mixture-of-experts system designed to be downloaded and run on your own hardware. Mistral’s pitch is that it is the strongest such option built outside China — and, notably, deployable on European infrastructure independent of US clouds. The company says Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in its own European datacentres.
  • On paper, the specs are in the same league. Press coverage of the launch reports DeepSeek’s current flagship at around 1.65T total parameters and Qwen’s at roughly 2.4T, versus Large 4’s 1.05T — but Large 4’s 52B active-parameter count is close to DeepSeek’s reported ~49B active parameters, and reported benchmark margins between the three are described as narrow, with no single model holding a decisive lead across the board. These are second-hand figures; the real comparison starts when independent labs can run all three side by side.
  • Cyber is Mistral’s chosen battleground. Neither DeepSeek nor Qwen has made cybersecurity the centerpiece of a flagship launch the way Mistral just did. If the 82% Cyber Index and 93% Cybench figures hold up under independent testing, that is a genuine differentiator for security teams — and a provocative one, given that the pre-release testing program hands a reduced-safety version to government testers first.

The honest bottom line on “Mistral Large 4 vs DeepSeek”: anyone declaring a winner today is guessing — and anyone crowning it the best open weight AI model of 2026 is guessing even harder. The weights aren’t public, the benchmark cards aren’t published, and the CEO’s comparison didn’t name names. Check back after October 27, when independent evaluators can actually run the model — we plan to update this article then.

What We Still Don’t Know

A responsible “everything we know” has to include the gaps:

  1. The weight license. “Open-weight” is not “open source.” Until Mistral publishes the license, we don’t know the commercial-use terms, attribution requirements, or any use restrictions.
  2. Full benchmark methodology. Company-reported scores without published test cards can’t be independently reproduced or stress-tested.
  3. Real-world behavior. Pre-release red-teaming with a reduced-safety version is reassuring process-wise, but the public won’t see those results — only the final, moderated model.
  4. Pricing beyond preview. API preview pricing is known; long-term pricing and the economics of self-hosting are not.
  5. The “beats Chinese models” claim. Unspecified models, unspecified benchmarks, no public evidence yet.

Bottom Line

Mistral Large 4 (“Le Chonk”) is the most significant European AI launch of 2026: a trillion-parameter, open-weight, natively multimodal model with a genuine emphasis on cybersecurity, arriving October 27 with weights to follow. The announcement-day facts are solid; the performance claims are promising but — as with every vendor launch — await independent verification.

If you’re evaluating AI models for your own work, keep in mind that the model landscape moves fast and today’s benchmark leader can be next month’s runner-up. For hands-on tool comparisons you can act on right now, see our tested roundups — including AI coding tools and AI presentation makers — where every pick went through the same real-world testing process.

Last updated October 7, 2026. We will revise this article when the open weights land and independent benchmarks appear.

Frequently asked questions

When is the Mistral Large 4 release date?

Mistral Large 4 was announced on October 6, 2026, with API preview access available now through Mistral Studio. According to the company, full public availability begins October 27, 2026, and the open model weights are expected by the end of October.

Is Mistral Large 4 open source?

Not exactly. Mistral describes Large 4 as open-weight, meaning the model weights will be downloadable and runnable on your own hardware. That is different from open source, which also covers training code and data. The weight license had not been published as of October 7, 2026.

How big is Mistral Large 4 (Le Chonk)?

Mistral describes it as a 1-trillion-parameter model; its technical documentation lists 1.05 trillion total parameters with 52 billion active parameters in a mixture-of-experts architecture, plus a 1.6-billion-parameter vision encoder and a 1-million-token context window.

How does Mistral Large 4 compare to DeepSeek?

Both are open-weight mixture-of-experts models. Mistral claims Large 4 is the best open-weight model from the US or Europe on aggregated benchmarks, and its CEO told Reuters it sits 'above the Chinese models on certain aspects, including cyber' — a company claim made without naming specific models or public benchmark cards. Independent head-to-head comparisons will have to wait until the weights are released.

How much does the Mistral Large 4 API cost?

Preview pricing reported at launch is $1.36 per million input tokens and $4.18 per million output tokens. Pricing for self-hosting the open weights is not yet known.

Can I download Mistral Large 4's weights now?

No. As of October 7, 2026, only hosted API preview access is available. Mistral says the open weights will follow by the end of October 2026.