Mistral Large 4 "Le Chonk" Explained (2026): Europe's 1-Trillion-Parameter Open-Weight Bet

Updated: October 2026 · 9 min read · By the CreatorSprout team

An honest note up front: Large 4 entered public preview on October 6, 2026 and we haven't run it hands-on yet. Everything below is built from Mistral's launch materials and independent reporting — all benchmark claims are Mistral's own until independent testers get the weights. No affiliate links in this article.

On October 6, 2026, France's Mistral AI ended a five-month drought of major releases and unveiled Mistral Large 4 — internally codenamed "Le Chonk" — a 1-trillion-parameter open-weight model the company calls the strongest non-Chinese open model ever built. It landed exactly one day after NVIDIA-backed Reflection AI's Beam, turning one quiet October week into the most aggressive open-weight showdown we've seen this year.

Here's the announcement in Mistral's own words: "Today, we're launching a public preview of Mistral Large 4. Unofficially ML4, very officially: le Chonk." What's behind the nickname — and what it actually means for you — is below.

What is Mistral Large 4?

Mistral Large 4 (ML4) is the Paris lab's new flagship: a sparse mixture-of-experts model that fuses instruction-following, reasoning, and agentic capability into one release, with native multimodal input (text and images together). Mistral describes it as a "hybrid instruct-and-reasoning MoE" — one model you can use as a normal chatbot, a reasoning engine, or the brain behind tool-calling agents.

The context matters as much as the model. Mistral's last flagship was five months old, the open-weight frontier in 2026 has been overwhelmingly led by Chinese labs (DeepSeek, Qwen, Kimi, Z.ai's GLM), and Mistral just closed a €3 billion Series D at a €21-billion-plus valuation in September — the largest equity round ever raised by a European tech company. CEO Artur Mensch framed the launch as a direct challenge at an event in Abu Dhabi: the claim that Europe cannot compete in AI "is not true." This model is the bet that backs that statement.

Specs at a glance

All figures below come from Mistral's launch materials and independent launch-day reporting:

The training stat is the one that made the industry blink: a frontier-scale trillion-parameter model trained from scratch in about two months on roughly 4,000 of NVIDIA's Grace Blackwell chips. That's an unusually fast cycle — Mistral says its research operation has scaled from three researchers to roughly 300.

Benchmarks: Mistral's claims vs reality check

Mistral's launch framing was unusually combative. The company positioned Large 4 directly against the Chinese open-weight field — DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, and GLM-5.3 — and claimed it is the strongest open-weight model developed outside China "by a substantial margin."

The specific claims, all company-reported:

• Coding: best among open models on SciCode-Verified (scientific coding) and on a combined coding-agent index blending several agentic-coding tests, beating DeepSeek V4 Pro and Qwen3.8 Max.
• Cybersecurity, finance, manufacturing: state-of-the-art among open weights on Mistral's reported critical-workload suites.
• Visual grounding: outperforms closed frontier models on visual grounding tasks (Mistral's claim — notable because open models usually lose this category).

The reality check: these are preview benchmarks published by the company that sells the model. The independent test only becomes possible when the weights are downloadable — expected late October. Until then, treat every ranking as marketing data, not measurement. That said, the comparison set is telling: Mistral chose to fight on coding and cyber, the categories where open models are most verifiable — and where anyone will be able to check its work within weeks.

Large 4 vs Reflection Beam: the one-day showdown

Reflection AI unveiled its 501B-parameter open-weight model Beam on October 5 — one day before Le Chonk. Both pitch themselves as the West's answer to Chinese open models. How they compare on published specs:

The honest read: Beam is the safer bet on paper right now — confirmed Apache 2.0 and lower inference cost. Large 4 is the bigger swing: twice the parameters, bolder benchmark claims, a cybersecurity story nobody else is telling. But a promised license is not a license, and a promised October 27 weights drop is not a download link. The ranking between these two gets decided when both can be run independently — likely early November.

Availability and pricing

What launched October 6 is a public preview through Mistral's API and Mistral Studio — available to all users, running on Mistral's own machines. According to Mistral's documentation, the preview supports chat completions, structured outputs, function calling, document Q&A, batching, and agents/conversations with built-in tools. Mistral also offers EU deployment operated end-to-end by Mistral under European law.

Pricing needs a flag. Mistral's launch materials list $1.36 per million input tokens and $4.18 per million output tokens. But the model's documentation page has shown different figures — $0.68 input, $0.07 cached input, $2.09 output per million tokens — displayed as lower values against struck-through higher ones. Independent launch-day coverage noted the same mismatch. Because the model is in preview and the two pages disagree, check the live pricing in Mistral Studio before budgeting any production workload — don't quote either number to your boss.

Pricing note: these figures come from Mistral's pages and launch-day reporting on October 6–7, 2026 and change often. Verify on Mistral's current pricing page before buying.

Open weights: what "open" actually means here

Large 4 is not yet downloadable. Mistral says the weights arrive at the end of October — Reuters reports October 27, 2026 as the target. Until that day, the open-weight story is a promise: the model runs only on Mistral's infrastructure.

Between now and the release, Mistral is running a staged testing period with developers, cybersecurity leaders, and government authorities. The license is still unannounced — Large 3 shipped under Apache 2.0, but Mistral said the Large 4 weights are "expected under a custom Mistral license," so don't assume Apache 2.0 carries over. That license choice is the single most important detail still to come: it decides whether businesses can actually self-host and fine-tune the model, or just rent it.

Mistral's pitch for open weights is worth quoting in spirit, because it explains who this model is for: co-founder and chief scientist Guillaume Lample argued that owning the model matters more than any single benchmark — with a closed model, he said, there's no guarantee it will still be there tomorrow. The enterprise and government buyer who wants a frontier model on their own infrastructure — or under European data law — is Large 4's target customer, not the casual chatbot user.

What this means for creators

A trillion-parameter open model isn't a creator tool the way ElevenLabs is — but it moves the ground under several creator workflows:

• Cheaper, better local AI assistants. When the weights ship, a model of this caliber becomes something developers can fine-tune and self-host. Expect better open-source coding assistants, script helpers, and research tools within months — tools you run without paying per-token API rent.
• EU-hosted inference for privacy-sensitive creators. Mistral's Europe-operated deployment under EU law is genuinely distinctive. Creators handling client data, unreleased content, or health/education audiences get a frontier-class option with a clearer compliance story.
• Stronger AI research for content production. Large 4's claimed strengths — coding agents, document Q&A, 1M-token context — map directly onto research-heavy content: feed it a 500-page report and get a script outline back. (We said "claimed" — verify once it's actually released.)
• The comparison shopping gets better for everyone. Two serious Western open models landing within 24 hours of each other (Beam and Large 4) puts real pressure on API pricing across the board. Even if you never touch either model, your ChatGPT and Claude bills benefit from the competition.

The honest catch: none of this is usable today in open form. The preview API is the only door in until late October, and the two competing prices make cost planning awkward. Creators should watch the weights release and the license — not the benchmark charts.

Pros and cons

The cybersecurity angle: read this first

This is the part of the launch most coverage skims, and it's the most consequential. Mistral is pitching Large 4 as a cyber-capable model — strong at both offensive security benchmarks and cyber defense — and it's handling that capability unusually carefully: before the weights ship, cybersecurity experts, vetted partners, and government authorities get access to a version with reduced safety moderation and expanded cyber capabilities for red-teaming.

What that tells you:

• This is a genuinely dual-use release. A model good at cyber defense is, by construction, good at finding the vulnerabilities first. Mistral's staged rollout is an admission of that.
• The reduced-moderation test version is the story. How Mistral gates that version — and what the red-teamers find — will shape regulation of open models well beyond this release.
• For creators and small businesses, the takeaway is defense. The same capability that finds holes can audit your stack: expect "AI security audit" tools built on models like this within months. If you run a business online, AI-assisted vulnerability scanning is about to get much better — and attackers get the same upgrade.

Our early verdict

FAQ

When did Mistral Large 4 launch?

October 6, 2026, in public preview through Mistral's API and Mistral Studio. The downloadable open weights are expected around October 27, 2026.

What does "Le Chonk" mean?

It's Mistral's internal nickname for Large 4 — a play on the model being enormous (1.05 trillion parameters). Mistral put it in the announcement itself: "Unofficially ML4, very officially: le Chonk."

Is Mistral Large 4 really open source?

Not yet. The preview runs only on Mistral's infrastructure. The weights are expected by end of October 2026, and the license hasn't been announced — it may be a custom Mistral license rather than Apache 2.0, so "open weights" and "open source" may not mean the same thing here.

How much does the Mistral Large 4 API cost?

Mistral's launch materials list $1.36 per million input tokens and $4.18 per million output tokens, but the model's documentation page has shown lower figures ($0.68 input, $0.07 cached input, $2.09 output). Check Mistral Studio's live pricing before budgeting.

Mistral Large 4 vs Reflection Beam — which is better?

Too early to say. Beam (launched Oct 5) has a confirmed Apache 2.0 license and lower inference cost (23B active params vs 49B). Large 4 has twice the parameters and bolder benchmark claims, but they're company-reported and the license is pending. Independent rankings come after the weights release.

Can I run Mistral Large 4 on my own hardware?

Not until the weights are released (~October 27, 2026), and self-hosting a trillion-parameter model needs serious GPU infrastructure — this is a data-center model, not a laptop model. Watch for quantized community versions after release.

Is Mistral Large 4 good for creators?

Indirectly but meaningfully: strong open models drive down API prices and power better open-source assistants, script helpers, and research tools. The EU-hosted option also gives privacy-sensitive creators a clearer compliance story. The direct benefits arrive once the weights ship.

Sources

• VentureBeat: Mistral debuts Large 4 'Le Chonk' (Oct 6, 2026)
• Reuters: France's Mistral launches AI model it says outperforms some Chinese rivals (Oct 6, 2026)
• Unite.AI: Mistral unveils 1.05T-parameter 'Le Chonk' in public preview
• Tech Insider: Mistral Large 4 spec sheet and positioning
• Digital Pulse Brief: pricing, benchmarks & open weights explained
• TestingCatalog: Mistral launches Large 4 preview with 1T parameters