How to Use Mistral Large 4 'Le Chonk' Preview (2026)

Last updated: October 13, 2026

Europe finally has a trillion-parameter model — and you can call it through an API today. Mistral AI launched Mistral Large 4, nicknamed “Le Chonk”, as a public preview on 6 October 2026, with the open weights promised for the end of the month. It is the biggest open model yet from a Western lab, it is natively multimodal, and Mistral is pricing the preview aggressively. But “open” does not mean downloadable yet, and the sale pricing has no published end date. This guide walks through what Large 4 is, how to try the preview right now, what it will actually cost you, and whether you should build on the preview or wait for the weights.

Quick answer: Mistral Large 4 'Le Chonk' in API preview: 1.05T-param MoE, 1M context, $0.68/1M tokens on sale. Costs, benchmarks, and whether to wait for the 27 Oct weights.

What is Mistral Large 4 ‘Le Chonk’?

Mistral Large 4 is Mistral AI’s new flagship general-purpose model, announced 6 October 2026 at a launch event in Abu Dhabi. It is a granular mixture-of-experts (MoE) model: instead of one giant network doing everything, it splits its parameters across many smaller “expert” networks and routes each token through only a few of them. The result is a model that stores a huge number of parameters but activates a small fraction per token.

The headline specs, per Mistral’s announcement and model card: 1.05 trillion total parameters, a natively multimodal 1.6-billion-parameter vision encoder, and a 1-million-token context window. CEO Artur Mensch framed the launch as a direct answer to the claim that Europe cannot compete on frontier AI, telling Reuters the model sits “above the Chinese models on certain aspects, including cyber”. Mistral trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacentres — an infrastructure-independence statement as much as a technical one. It handles more than 160 languages, including every official language of the EU.

One honest caveat on the specs: Mistral’s own pages disagree on how many parameters are active per token. The announcement says 49 billion; the model card says 52 billion. Until Mistral corrects one of them, treat the active count as somewhere between 49 and 52 billion — the difference is rounding-level for planning purposes, but it is worth knowing before you quote a number.

How to try the preview right now

The preview is live through Mistral’s API today; the only thing missing is the downloadable weights. Here is the shortest path from zero to your first call.

Create a Mistral Studio account and generate an API key. Sign up at Mistral’s platform, open the API keys section, and create a key. Keep it server-side — the preview is a paid API, so treat the key like a credential.
Set the model identifier. The preview model id is mistral-large-4. Mistral’s own cookbook notebook for the launch — mistral/image_understanding/bbox_getstarted.ipynb in the mistralai/cookbook GitHub repository — sets exactly this value, which confirms the id is live in production endpoints.
Send a first chat completion. Point your existing Mistral-compatible client at the chat completions endpoint with the new model id. If you already use Mistral’s API, the only change is the model string.
Try the vision side. Large 4 is natively multimodal, so pass an image in the message payload rather than calling a separate vision endpoint. Mistral’s launch-day cookbook adds a bounding-box and object-detection notebook, which is a good first vision test.
Turn on prompt caching for anything repeated. Cached input reads cost $0.14 per million tokens at list price ($0.07 on the current sale) — roughly a tenth of fresh input. For agent loops and RAG pipelines that resend system prompts, caching is where the MoE pricing really starts to bite in your favour.

A practical note: this guide is desk-researched, not hands-on tested. None of the steps above were run against a live key for this article — verify the model id and endpoint behaviour in Mistral’s docs before wiring it into production.

What it costs: list pricing vs the launch sale

Mistral lists the preview at $1.36 per million input tokens, $0.14 per million cached-input tokens, and $4.18 per million output tokens. But the API is currently on sale — and the sale price is the one you will actually pay today.

Price tier	Input / 1M tokens	Cached input / 1M	Output / 1M tokens
List price	$1.36	$0.14	$4.18
Current sale	$0.68	$0.07	$2.09

The sale figures are reported by Developers Digest, which notes Mistral has not published an end date for the sale, and that Vercel AI Gateway passes the same sale price through. Treat the sale as temporary until Mistral says otherwise — build your cost model on list price, and treat the sale as a bonus.

Some desk arithmetic on what that means per task. Take a typical heavy workload: 1 million input tokens and 250,000 output tokens (roughly a large repo scan with a long write-up). At sale pricing that is $0.68 + $0.52 = $1.20. At list pricing it is $1.36 + $1.05 = $2.41. For comparison, the same workload on a Claude Opus-tier model at $5/$25 per million tokens would run about $11.25. Le Chonk’s MoE design is the whole reason it can sit in a mid-tier price band with a trillion-parameter body: only ~5% of the weights fire per token, so you are billed for the 52 billion that wake up, not the trillion that sleep. (Arithmetic is the author’s own, calculated from the published prices above; not hands-on measured.)

How good is it? Vendor claims vs independent tests

This is where the honest split matters. Mistral’s launch materials lean on the model’s strong suits; independent measurements paint a more mixed picture.

Claim	Status	Source
1.05T total params, 52B active, 1M context	Confirmed (vendor; active count 49B vs 52B on Mistral’s own pages)	Mistral announcement + model card, via techaiwire
Open weights “by the end of October”	Stated by vendor; Reuters reports the public release date as 27 October	Mistral announcement; Reuters via Rappler
Security partners and state authorities get an early, less-restricted version	Reported	Reuters via Rappler, 6 Oct 2026
93% on Cybench, 82% on CyberGym-E2E	Vendor-reported	MarkTechPost, 6 Oct 2026
Intelligence Index 38.38 (vs GLM-5.3 Max 44.78)	Independent, read 8 Oct 2026	The AI Realist, citing Artificial Analysis
33rd of 45 on the Vals Index	Independent	The AI Realist, citing vals.ai
Vuln-reproduction 81.7–82%, but 32nd of 44 on aggregate benchmarks	Independent	9 Oct independent tests
Launch sale has no published end date	Confirmed absent from Mistral’s pricing page	Developers Digest

The pattern is clear: Large 4 is genuinely strong in cybersecurity — Mistral’s own numbers and the independent vuln-reproduction scores agree — but on broad general benchmarks it sits mid-pack. One developer’s custom-language coding benchmark found it roughly level with Kimi K3 and Claude Sonnet 5.5 on that task, but slower than Sonnet 5.5 and more expensive than both (reported by Developers Digest). And Artificial Analysis still flags the model as proprietary, not open, until the weights actually ship. The fair summary: a strong contender in its best domains, not a champion across the board.

Try the preview now, or wait for the 27 October weights?

This is the decision most readers actually face. It depends on what you need.

Your situation	Verdict
You want to evaluate it for a product decision	Try the preview now — the API is live and the sale pricing makes evaluation cheap
You need EU-sovereign inference or on-premise deployment	Wait for the 27 October weights — the preview is API-only and Mistral-cloud only
You want the cheapest possible inference long-term	Wait — self-hosting the weights changes the cost equation entirely, and the preview sale may end
You build security tooling	Try now — cyber is its strongest verified domain, and security partners get early access anyway
You need the final checkpoint	Wait — co-founder Guillaume Lample said a final version ships with the weights, so the downloadable checkpoint may differ from today’s preview

The middle path is the most sensible for most teams: run your evaluation workload on the preview this week, and make the build-or-buy call after the weights and the final licence terms land. Mistral’s open-weight licence was still marked “coming soon” on launch day — do not assume the licence will suit commercial redistribution until it is published.

Limitations to know before building on it

The licence is not published yet. “Open weights” is a promise, not a document. Commercial teams should wait for the licence text.
The preview may not be the final model. Lample’s statement that a final version ships with the weights means behaviour you measure today could change on 27 October. Do not lock benchmark conclusions to the preview.
Mistral’s own spec pages disagree. The 49B vs 52B active-parameter discrepancy is minor, but it signals that the documentation is still settling. Verify specs in the docs, not in launch-day coverage.
It is mid-pack on general benchmarks. If your workload is general coding or reasoning rather than security, the independent scores (33rd of 45, 32nd of 44) suggest testing head-to-head against Kimi K3 or Sonnet-class models before committing.
The sale will end. No end date is published. Any cost model that depends on $0.68/$2.09 pricing should have a list-price fallback.

FAQ

Is Mistral Large 4 really open source?

Not yet, and possibly never in the strict sense. The weights are promised by 27 October 2026, but the licence was marked “coming soon” at launch. Until both ship, treat it as a proprietary API preview with an open-weight roadmap — which is exactly how Artificial Analysis classifies it today.

How do I access the Mistral Large 4 preview?

Through Mistral’s API with the model id mistral-large-4. Create a Mistral Studio account, generate an API key, and call the chat completions endpoint. The weights for self-hosting arrive with the 27 October release.

How much does Mistral Large 4 cost?

List price is $1.36 per million input tokens and $4.18 per million output tokens ($0.14 cached). It is currently on sale at $0.68 input / $2.09 output, with no published end date for the sale. Vercel AI Gateway passes the sale price through.

Is Le Chonk better than Claude or GPT models?

In cybersecurity, the independent scores say it is genuinely competitive. On broad general benchmarks it ranks mid-pack (low 30s of ~45 models on independent leaderboards read 8–9 Oct 2026), below the top closed frontier models. It is best evaluated as a cost-efficient open-weight contender, not a frontier-beater.

Should I wait for the open weights?

If you need self-hosting, EU-sovereign deployment, or the final checkpoint, yes — the public release is dated 27 October, and co-founder Guillaume Lample said the downloadable checkpoint may differ from today’s preview. For evaluation, the API preview is live and cheap right now.

Sources and methodology

This article is desk research compiled 13 October 2026 from Mistral’s official announcement, Mistral’s model documentation, Reuters reporting on the launch, Developers Digest’s pricing and spec analysis, The AI Realist’s independent-benchmark roundup, techaiwire’s spec comparison, and MarkTechPost’s launch coverage. Benchmark figures are labelled vendor-reported or independent as shown; sale pricing is reported by Developers Digest with no end date published by Mistral. Nothing in this article was hands-on tested against the live API. Costs per task are the author’s own arithmetic from published prices.

Mistral AI, “Introducing Mistral Large 4” (official announcement): https://mistral.ai/news/mistral-large-4/ (opens in new tab)
Mistral Docs, Mistral Large 4 model page: https://docs.mistral.ai/models/mistral-large-4-0 (opens in new tab)
Reuters via Rappler, “France’s Mistral launches AI model”, 6 Oct 2026: https://www.rappler.com/technology/mistral-large-4-ai-model-launch/ (opens in new tab)
Developers Digest, “Mistral Large 4: Specs, Pricing, Benchmarks and Weights Date”: https://www.developersdigest.tech/blog/mistral-large-4-open-weight-model-2026 (opens in new tab)

Govind Dheda

Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!

How to Use Mistral Large 4 ‘Le Chonk’ Preview (2026)