Mistral just announced the biggest AI model it has ever built. Mistral Large 4 — nicknamed Le Chonk inside the company — is a 1 trillion-parameter multimodal model that will be released as open weights at the end of October 2026. It is available right now in public preview through Mistral’s API, it is priced aggressively, and it is aimed squarely at enterprises and governments that want to run a frontier-class model on their own infrastructure.
The short version: on October 6, 2026, the French AI lab Mistral unveiled Mistral Large 4 (ML4), a mixture-of-experts model with 1.05 trillion total parameters but only 49 billion active per token, native image input, and a one-million-token context window. The API preview costs $1.36 per million input tokens and $4.18 per million output tokens, and the open weights are expected on October 27. Mistral claims it significantly outperforms any open-weight model developed in the US or Europe, with its strongest results in cybersecurity — a domain where several closed frontier models refuse to participate at all.
What is Mistral Large 4?
Mistral Large 4 is a general-purpose multimodal AI model built on a mixture-of-experts (MoE) architecture. The headline number is 1.05 trillion total parameters, but that figure needs a qualifier: only 49 billion of those parameters activate for any given token. In practice, each query costs about what a 49-billion-parameter dense model would cost to run — though you still need enough memory to hold the full model somewhere before the routing math can happen.
The model is natively multimodal: a 1.6-billion-parameter vision encoder gives it image input without a bolt-on pipeline, and it accepts a context window of up to one million tokens. Mistral says it trained across more than 160 languages, including every official language of the European Union. It is a hybrid model in a broader sense too — one system that handles instruction following, reasoning, and agentic operation, rather than splitting those jobs across separate checkpoints.
On Mistral’s own site, the company calls this one “Unofficially ML4, very officially: le Chonk” — a joke about the model’s size that has already become the story’s headline name.
The specs that actually matter
Parameter counts make good headlines but poor buying decisions. Here is what matters in practice:
- Architecture: granular mixture-of-experts, 1.05 trillion total parameters, ~49 billion active per token.
- Vision: 1.6-billion-parameter native vision encoder — image understanding is built in, not attached afterward.
- Context: one million tokens, putting it in the same league as the big frontier models for long-document work.
- Languages: 160+, covering every official EU language, which matters for the European enterprise and public-sector buyers Mistral is courting.
- Training: built from scratch over roughly two months on 3,800 Nvidia Grace Blackwell GPUs inside Mistral’s own European datacenters — not rented cloud capacity.
That last point is doing real work in the pitch. Training on its own hardware, in its own European datacenters, is Mistral’s argument for infrastructure independence: Europe can build and run frontier models without depending on American hyperscalers. For governments and regulated industries, where the data lives is as important as what the model can do.
Why cybersecurity is the wedge
Mistral is not pitching Large 4 as a general-purpose chatbot first. The company’s benchmarking materials lead with cybersecurity, and the numbers are striking: 93% on Cybench and 82% on a vulnerability reproduction and patching test.
The more provocative claim is about who does not play. Mistral notes that several leading closed models — including Claude Opus 5.5 and GPT-6 Astra — score near zero on the same vulnerability-reproduction test because they refuse to perform the task. Safety refusals are doing exactly what they were designed to do, but they leave a gap: security teams that need to reproduce vulnerabilities and test patches cannot use a model that declines. An open-weight model has no refusal layer once you run it yourself, so it fills that gap by construction.
Beyond security, Mistral claims state-of-the-art results among open models on coding, agentic workflows, finance and law workloads, and says its visual grounding — the ability to locate and describe specific elements in images — surpasses even some frontier closed models. It benchmarks Large 4 head-to-head against other open-weight systems: DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3 and GLM-5.3. The framing is deliberate: this is a race among open models, and Mistral wants to be leading it from Europe rather than catching up to China or the US.
On the widely watched DeepSWE coding benchmark, though, independent observers note Large 4 scores around 62% — solid on the open-weight leaderboard, but behind the closed systems clustered around 74%. The honest read is that open weights get you remarkably close to the frontier, not that they beat it.
Price, API access and how to try it today
You do not have to wait for the weights to use Large 4. A public preview went live on October 6 through Mistral’s API, and the pricing is set to undercut the big closed models: $1.36 per million input tokens and $4.18 per million output tokens. Mistral is also running a playground where you can try the model without writing any code.
The API model identifier is mistral-large-4. The preview is a paid, staged rollout: Mistral is giving developers, cybersecurity teams and government authorities roughly three weeks of access before the weights go public. The company says it will use that window to keep training with reinforcement learning and tune the final checkpoint that ships with the weights.
For context on what that pricing means in practice: it sits roughly in the mid-range of frontier API pricing — not the cheapest open-weight option, but far below what the top closed models charge. The MoE architecture is exactly what makes that pricing possible; the company only spends compute on the 49 billion parameters that matter for each token.
When the open weights actually land
This is the part that matters most, and it is also the part still in the future. The weights — the actual files that let anyone download, self-host and fine-tune the model — are expected at the end of October 2026, with October 27 reported as the target date.
Two caveats. First, until that date arrives, the open-weight release is a promise, not a product: the model runs only on Mistral’s own machines. Second, the license is not yet confirmed. Mistral’s previous flagship, Large 3, shipped under Apache 2.0; VentureBeat reported that Large 4 will use a custom Mistral license instead, which could restrict some commercial uses. The details arrive with the weights.
Once the weights are out, the pitch is straightforward: enterprises and governments can run and customize Large 4 themselves — including on sovereign infrastructure, with zero-data-retention options. That is a materially different proposition from an API, and it is the one that makes Mistral interesting to buyers who cannot or will not send their data to someone else’s cloud.
Why this release matters beyond the parameter count
Mistral’s chief scientist and co-founder, Guillaume Lample, has framed the open-weights argument in business terms rather than ideological ones: what really matters is owning the model, because a closed model could change or disappear tomorrow. If you build a product on someone else’s API, your margins and your availability are decisions somebody else makes.
The timing is also worth noting. Large 4 is the first major release funded by Mistral’s €3 billion Series D, which closed in September 2026 with Samsung leading — described as the largest equity round ever raised by a European technology company, at a valuation above €21 billion. The company says it already serves more than 125 large enterprises, including Airbus, ASML and HSBC. This is not a research demo from a startup; it is a product launch from a company with real enterprise customers and the capital to keep training.
It also continues a pattern worth watching. Open-weight models are compressing the gap to the closed frontier faster than many expected — from DeepSeek’s releases to Meta’s Llama family to now Mistral’s trillion-parameter play. The debate about the best AI assistant for your business is no longer just about which subscription to buy, as we covered in our Claude vs ChatGPT comparison for small businesses — it is increasingly about whether you need to buy a subscription at all, or just run the model yourself.
How Large 4 fits into the agentic AI wave
Large 4 arrives as AI agents are moving from demos into production. Mistral is pitching it explicitly for agentic workflows, and that matters because the agent layer is where a lot of 2026’s infrastructure bets are landing — from Cloudflare’s open-weight Clef decision models to always-on systems like OpenAI Dots and Meta Muse for small business.
An open-weight model that can run inside your own environment is a natural fit for agents that need access to internal data — exactly the workloads where sending everything to a third-party API is a non-starter for compliance reasons. If Mistral’s agentic benchmarks hold up, Large 4 becomes the open counterweight to the closed agent stacks the big labs are selling.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 (ML4), nicknamed Le Chonk, is a 1-trillion-parameter multimodal AI model announced by the French lab Mistral on October 6, 2026. It uses a mixture-of-experts architecture with 49 billion active parameters per token, supports image input and a one-million-token context window, and will be released as open weights at the end of October.
Is Mistral Large 4 free?
The API preview is paid: $1.36 per million input tokens and $4.18 per million output tokens. The open weights, expected around October 27, 2026, will be downloadable under a Mistral license — free to run on your own hardware, though the license terms are not yet confirmed.
When can I download the Mistral Large 4 weights?
Mistral says the weights ship at the end of October 2026, with October 27 reported as the target. Until then, the model is only available through Mistral’s API and playground. Note that a 1.05-trillion-parameter model needs substantial hardware to self-host; the 49-billion active parameter count sets the per-query compute cost, but the full model still has to fit in memory.
What does Le Chonk mean?
It is an internet-slang word for a large, chunky animal — Mistral’s own joke about the model’s size. The company introduced the nickname in its announcement materials: unofficially ML4, very officially Le Chonk.
How does Mistral Large 4 compare to GPT-6, Claude and Gemini?
Mistral positions it against other open-weight models (DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, GLM-5.3), claiming it significantly outperforms any open-weight model from the US or Europe. On coding benchmarks like DeepSWE, it trails the closed frontier systems — around 62% versus roughly 74% — but its standout results are in cybersecurity, where closed models that refuse the task score near zero.
Who should use Mistral Large 4?
Security teams that need vulnerability reproduction and patching without refusal layers, enterprises that want a frontier-class model on their own or sovereign infrastructure, European public-sector organizations that need EU-language coverage and data residency, and developers who want a powerful open model to fine-tune. For everyday chat and productivity, the closed assistants in our Claude vs ChatGPT comparison are still simpler to adopt.