Mistral Large 4: Pricing, Access, and Open-Weight Release Date
Mistral Large 4 is live in API preview on Mistral Studio now, with open weights promised by month's end. Here's how access works.
What is Mistral Large 4?
Mistral Large 4, nicknamed “Le Chonk” internally, is Mistral AI’s newest large language model, built as a mixture-of-experts system with roughly a trillion total parameters and about 50 billion active at any given time. It was trained in Mistral’s own European data centers and is currently available as a public preview rather than a finished, generally available product. The company is rolling it out in two stages: an API you can use right now, and open-weight files that are expected to follow within the same month.
TL;DR
- Public preview status means Mistral Large 4 is usable today through the API, but the model is still being positioned as a preview release rather than a final, stable version.
- Mistral Studio is the current access point for developers who want to call the model programmatically before the open weights ship.
- Open weights are promised by the end of the month, which will let anyone download and self-host the model instead of relying solely on the hosted API.
- The model mixes mixture-of-experts efficiency with scale, running at about 50 billion active parameters out of a total near one trillion, which keeps inference faster than a dense model of similar size.
- Early third-party testing shows strong coding and security benchmarks, with Mistral Large 4 beating most open models on tasks like finding vulnerabilities in code, though some Chinese open models still lead on specific benchmarks like terminal command handling.
- Multilingual performance is uneven, holding up well on major world languages but degrading noticeably on low-resource languages and scripts.
- No confirmed public pricing was available at launch, so budgeting for production use will depend on what Mistral publishes once the model exits preview.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How do you access Mistral Large 4 right now?
The only confirmed access path at launch is the API, hosted through Mistral Studio. Developers who want to test the model’s coding, reasoning, or multimodal capabilities can call it there during the preview window. This mirrors how most large model vendors stage a release: ship the API first so developers and enterprise customers can start integration and evaluation work, then follow with open weights for people who want local or self-hosted deployment.
There’s no indication in the current preview of a separate consumer chat surface being the primary access point. For builders, Mistral Studio is the practical entry point today. If you’re testing the model for coding assistants, security tooling, or document analysis pipelines, this is where that work happens until the weights are public.
When are the open weights actually coming out?
Mistral has committed to releasing the open weights by the end of the same month as the preview launch. That’s a notably short gap between API access and open release, a pattern Mistral has followed before, favoring openness as a differentiator against closed-weight competitors. Once the weights land, developers will be able to download and run the model on their own infrastructure, which matters for teams with data residency requirements, custom fine-tuning needs, or a preference for avoiding per-token API costs at scale.
No exact calendar date has been published beyond “later this month,” so teams planning around the open-weight release should treat that as a near-term but not fixed deadline.
What does “public preview” mean for production use?
Calling a model a public preview is a signal that it’s stable enough for real testing but may still see adjustments before a full release. Pricing, rate limits, and even some model behavior could shift between the preview phase and general availability. For teams evaluating Mistral Large 4 for production systems, this is worth treating as a research and prototyping window rather than a green light for mission-critical deployment. The upside is that you get hands-on access to a frontier-scale model early, which is valuable for benchmarking against incumbents like GPT and Claude-class models, or against open-weight rivals.
How does Mistral Large 4 perform on real tasks?
Independent testing (notably from AI YouTuber Fahad Mirza, who ran the model through a battery of practical tasks) found a mixed but largely positive picture.
On a benchmark measuring real software engineering fixes, Mistral Large 4 outperformed most open models tested, trailing only Kimi K3. On a terminal-command handling benchmark, it did well against most rivals but was clearly behind GLM 5.3. On a cybersecurity benchmark measuring the ability to find and fix security flaws in real code, it tied for the top spot with GLM 5.3 Flash. On a benchmark testing multi-app business workflows (email, spreadsheets, Slack-style tasks), it landed just behind GLM 5.3 among open models. On a financial research and analysis benchmark, it sat near the top, just behind GLM and slightly ahead of a GPT-6 class model referenced in the chart.
In hands-on testing, the model also performed well on a planted security vulnerability: given a Flask-based client portal with a deliberately broken access control bug (an insecure direct object reference, or IDOR), it correctly identified that a delete endpoint checked user ownership while a get endpoint didn’t, exposing one user’s invoice data to another logged-in user. It also flagged additional issues beyond the planted bug, including plaintext password storage, a hardcoded session secret, and a leaked session token in a committed file, and it proposed working code fixes.
On a scientific reasoning task involving electrical shock scenarios (reading a diagram and working out which of several people touching wires would actually be shocked), the model reasoned through the setup correctly, including a trick case involving someone touching both hot and neutral wires.
It also handled a social reasoning task well, correctly interpreting an ambiguous WhatsApp exchange with mismatched timestamps, picking up on workplace slang and figuring out how a message got misread in a personal context.
Where it struggled was multilingual translation. Asked to translate a headline into around 75 languages, it handled major world languages (Mandarin, Arabic, Japanese, Thai) cleanly but broke down on low-resource languages like Gujarati.
Is Mistral Large 4 worth testing now?
For developers who want early access to a large, mixture-of-experts model with strong coding and security benchmark results, yes, worth testing through the API now, especially if your use case touches code review, vulnerability detection, or financial document analysis, areas where it scored well in both formal benchmarks and hands-on testing. If your priority is low-resource language support, the preview suggests you should wait and verify performance directly rather than assume parity with major-language results.
For teams specifically waiting on self-hosting, the open-weight release later this month is the more relevant milestone. It removes dependency on the hosted API and opens the door to fine-tuning, local inference, and cost control that a pure API relationship doesn’t offer.
Frequently Asked Questions
What does “Le Chonk” mean?
It’s an internal nickname Mistral used in its own announcement, playing on internet slang for something large and chunky. The model page reportedly includes a pixel-art cat graphic to match the name. The official model name is Mistral Large 4.
How big is Mistral Large 4?
It’s a mixture-of-experts model with roughly one trillion total parameters and about 50 billion active parameters during inference, which keeps compute costs lower than running a dense trillion-parameter model directly.
Where can I use Mistral Large 4 today?
Through the API on Mistral Studio, which is live now as part of the public preview. Open weights for self-hosting are expected by the end of the same month.
Is Mistral Large 4 pricing confirmed?
No confirmed public pricing details were available at the time of the preview launch. Anyone budgeting for API use should check Mistral’s own documentation directly before committing to production usage.
How does it compare to other open models?
It’s competitive with or ahead of most open models on coding and cybersecurity benchmarks, tying for the top spot on a security vulnerability benchmark. It trails models like GLM 5.3 on terminal command and business workflow tasks, and Kimi K3 on software engineering fixes, so it’s strong but not uniformly dominant.