Mistral Large 4: Europe’s Trillion-Parameter Open Model, Read Before You Host

Mistral Large 4 is a trillion-parameter open-weight model built in Europe. What the licence allows, what hardware it needs and where it beats rivals.

Disclosure: this guide is based on Mistral’s announcements, the model card and independent coverage, including early Artificial Analysis results. We have not self-hosted Mistral Large 4, which needs an eight-GPU class server. Figures are attributed in the text.

Mistral released the open weights of Mistral Large 4 on October 29, 2026, three weeks after an API-only preview. The model is nicknamed Le Chonk, a cat meme the company leaned into, and it is the largest open release Mistral has ever made: roughly one trillion parameters, 49 billion active per token, a one-million-token context and native multimodal input.

The honest summary: Mistral Large 4 is the strongest frontier-scale open model trained outside China and the United States, and its coding and security focus is real. It is not at the level of the best closed flagships, and early independent testing from Artificial Analysis said so plainly. Its licence also has a clause about offering the model as a service that you need to read before building a hosted product.

Table of Contents

Mistral Large 4 steel bridge engineering structure
Mistral Large 4 is a bridge, not a peak: it trails the closed flagships but gives Europe a frontier-scale model it can run itself.

What Mistral Large 4 Is

Mistral Large 4 is a mixture-of-experts model. Reports put the total at about one trillion parameters, with Tech Insider citing 1.05 trillion. Active parameters are 49 billion per token, or 52 billion counting embeddings, which explains why some outlets quote one figure and some the other.

It is natively multimodal and multilingual, and Mistral says it was trained from scratch on European infrastructure. The context window is one million tokens. It offers three reasoning modes, Low, Medium and High, with High as the default.

The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, in FP8 and NVFP4 formats. Shipping a 4-bit NVFP4 build alongside FP8 is a practical choice: it roughly halves the memory needed compared with FP8, at a small quality cost.

Mistral positions the model for agentic coding and cybersecurity, specifically defensive security work. That focus shows in the benchmarks it chose to highlight.

What Changed Since the Preview

Mistral opened an API-only public preview on October 6. The open weights followed 23 days later. In between, the company says it continued scaling reinforcement learning and incorporated feedback from customers and partners.

The measurable result: AI TLDR reported a Terminal-Bench 4.0 score of 41%, up 13 points from the preview. SiliconANGLE described the release as having received three additional weeks of reinforcement training and customer testing. That is a large jump for three weeks, and it suggests the preview was released before post-training had finished.

For anyone who tested the preview and was unimpressed, this matters. The released weights are a meaningfully better model than what the API served in early October. Retest before writing it off.

The Mistral Large 4 Licence

Mistral Large 4 ships under the Mistral Ecosystem Open-Weight licence, which AI TLDR and others describe as a modified MIT licence. For most uses it behaves like MIT: you can download, run, modify and fine-tune it.

The modification is a model-as-a-service clause. Braindetox’s analysis singled it out as the part to read. The clause concerns offering the model as a hosted service to third parties, which is exactly what a cloud provider or an AI startup reselling inference would do. Internal use and building products on top of the model are a different matter from selling raw access to it.

Compsmag’s summary was direct: developers should read the licence before building a commercial service around the release. That is good advice for every custom open licence, and it applies here.

Compared with its peers, the terms sit in the middle. DeepSeek V4.1 Flash uses plain MIT with no service clause. Meta’s Muse Spark Open requires a visible attribution in every product. NVIDIA’s Nemotron 3 Ultra carves outputs out of its licence. Mistral’s clause targets a narrower group, mainly people who would compete with Mistral’s own API.

Mistral Large 4 Benchmarks

Benchmark Score Context
Terminal-Bench 4.0 41% Up 13 points from the preview
FrontierCode 2.0 (Cognition, private) 19.7% Mistral says first among open models
Artificial Analysis Coding Agent Index 48% Independent
Artificial Analysis Cyber Index 51% Independent

Mistral claims the model leads open-weight models on FrontierCode 2.0 and Terminal-Bench 4.0. Kie.ai reported about 47% fewer tokens per coding task than GLM-5.3, which matters for cost and agent latency, and a 96.6% refusal rate on a set of malicious requests, relevant given the security focus.

The independent picture is more sober. The Register reported that early Artificial Analysis testing put Le Chonk at a major disadvantage relative to OpenAI’s and Anthropic’s flagship models. That is not a contradiction. Leading the open field and trailing the closed frontier can both be true, and for this model they are.

Hardware for Self-Hosting Mistral Large 4

A trillion parameters is a lot of memory. Simple arithmetic: FP8 needs about one byte per parameter, so roughly 1 TB for weights. NVFP4 needs about half a byte, so roughly 500 GB. Add the context cache on top, which grows with the length of each conversation.

Braindetox noted that Mistral’s documentation includes eight-GPU deployment examples. With current high-memory data-centre GPUs, an eight-GPU node can hold the NVFP4 weights with room for a working cache. FP8 needs more headroom.

The active parameter count, 49 billion, keeps per-token compute manageable once the weights are loaded. Throughput depends on batching: a node serving many concurrent users spreads the fixed memory cost across more requests.

Compsmag’s point is worth repeating: open weights do not remove deployment constraints. They give you control over where and how the model runs. You still need the hardware, the operations skills and the patience to tune an inference stack for a model this size.

Mistral Large 4 Through the API

For most teams, Mistral’s hosted API is the practical way in. Pricing is $0.70 per million input tokens and $2.10 per million output tokens, according to AI TLDR.

That places it well below closed flagships. Claude Opus 5.5 lists at $4 and $20, GPT-6.1 Sol at $2 and $10. It is above the cheapest open options, such as DeepSeek V4.1 Flash at $0.15 off-peak, but those are smaller-active models with different strengths.

The reasoning modes give you a cost dial. High is the default and the most expensive in output tokens. Medium is a sensible setting for everyday coding help. Low suits extraction and classification, where heavy reasoning adds cost without improving the result.

How It Compares With Other Open Giants

Model Total / active Licence Notable
Mistral Large 4 About 1T / 49B Modified MIT, service clause European, coding and cyber focus
Qwen3.8 Max 2.4T / 95B Open weights Largest on paper
DeepSeek V4 Pro 1.6T / 49B Open weights Traffic now routed to V4.1 Flash
Muse Spark Open 816B / 42B Muse Model License Most capable US open model
Nemotron 3 Ultra 550B / 55B OpenMDW-1.1 Native NVFP4, open data

Tech Insider’s comparison put Qwen3.8 Max ahead on raw size. Size is not capability, but it shows Mistral is not trying to win a parameter race. Its differentiators are origin, licence and a specific focus on coding and defensive security.

The token efficiency claim deserves attention. If Mistral Large 4 really uses about half the tokens per coding task of a rival, its effective cost per task is lower than its per-token price suggests. That is the kind of figure to verify on your own tasks, because it varies with workload.

The Sovereignty Argument

Arthur Mensch, Mistral’s co-founder and chief executive, framed the open release as a counterweight to concentrated computing power. That argument resonates in Europe, where many governments and companies want frontier AI they can run under their own control.

Korben’s coverage summarised Mistral’s claim as the best open model developed outside China. For a European public body or a company with data-residency rules, a model trained in Europe, downloadable and hostable in its own data centre, solves problems that a US-hosted API cannot.

AI Market Watch observed that the release adds a European option to a market where Chinese labs are prominent suppliers of advanced open-weight models. For buyers who prefer not to depend on either US cloud APIs or Chinese-developed models, Mistral Large 4 is one of very few frontier-scale choices.

The trade-off is capability. Choosing sovereignty means accepting a model that trails the closed frontier. For many regulated workloads, that is an easy trade. For teams competing on raw coding ability, it is a harder one.

Mistral Large 4 for Defensive Security Work

Mistral chose to highlight cybersecurity alongside coding, and that is an unusual emphasis for an open release. Security teams have specific needs that make an open model attractive: they often work with sensitive logs, malware samples and internal vulnerability data that cannot be sent to a third-party API.

A self-hosted Mistral Large 4 can triage alerts, summarise incident timelines, explain suspicious scripts and draft detection rules without that data leaving the network. A 51% score on Artificial Analysis’s Cyber Index places it among the stronger open options for this work.

The refusal behaviour matters here too. Kie.ai reported a 96.6% refusal rate on a set of malicious requests. For a model that will be available to anyone who downloads it, strong refusals on clearly harmful requests are part of what made the release acceptable. For defenders, the risk is the reverse: over-refusal on legitimate analysis. Test with realistic defensive tasks before relying on it.

As with any security tool, keep a human in the loop. A model that explains a script can be wrong about what it does, and a detection rule drafted by a model needs testing before it goes into production.

Why the Nickname Matters More Than It Should

It is easy to roll your eyes at a frontier model called Le Chonk. The nickname comes from an internet cat meme about very large, very round cats, and Mistral leaned into it in both announcements and the model card. The Register called the official product name a lot less fun than the very official moniker.

There is a serious point underneath. Model naming across the industry has become so confusing that a memorable nickname actually helps people find the right thing. Developers searching for the release used the nickname at least as often as the product name, which is why this guide uses both. If you are writing internal documentation, record the full model ID as well, because nicknames do not survive version changes and the next large release may need a new joke.

Fine-Tuning and Distilling Mistral Large 4

Few teams will fine-tune a trillion-parameter model directly. The more common pattern with Mistral Large 4 is distillation: use the large model to generate high-quality examples for a narrow task, then train a much smaller model on those examples and deploy the small one.

This works well for tasks such as classifying security alerts, extracting fields from contracts or converting legacy code patterns. The large model supplies careful, consistent labels; the small model learns to reproduce them at a fraction of the serving cost.

Check the licence before you start. Licences differ on whether outputs and derived models carry obligations, and Mistral’s terms include a clause about hosted services. Confirm that your distilled model, and how you plan to offer it, fit within the terms.

If you do fine-tune Mistral Large 4 itself, adapter methods that train a small set of extra weights are the realistic route. They fit on the same hardware you use for inference and produce small files that are easy to version and swap.

A Sensible Rollout Plan for Mistral Large 4

Start with the hosted API for two to four weeks. That tells you whether Mistral Large 4 handles your tasks well enough before you commit to buying or renting an eight-GPU server.

Measure cost per completed task, not just per token. The reported token efficiency on coding work could make the effective cost lower than the headline rate, but only your workload will tell you by how much.

If results are good and you need data control, move to self-hosting with the NVFP4 weights first. They need roughly half the memory of FP8, and for most workloads the quality difference is small. Keep the API as a fallback while you tune the self-hosted stack.

Finally, schedule a review in three months. Open models are improving quickly, and Mistral will almost certainly ship updates. A short, repeatable evaluation set makes each review an afternoon rather than a project.

Who Should Use What

If you are Start with Why Link
A European organisation needing sovereign AI Mistral Large 4, self-hosted Trained in Europe, open weights Read the release
A security team with sensitive data Mistral Large 4 on your own hardware Cyber focus, data stays local See specs
Wanting the best coding at any price A closed flagship such as Claude Opus 5.5 Independent tests show a gap See Opus 5.5
Wanting the cheapest open API DeepSeek V4.1 Flash Lower price, MIT licence See specs

Related on this site: for more context, see Muse Spark Open, Claude Fable 5.2 vs Mistral Large 4 vs Kimi K3.1 Pro and Nemotron 3 Ultra.

Frequently Asked Questions About Mistral Large 4

What is Mistral Large 4?
Mistral’s largest open-weight model, nicknamed Le Chonk: a mixture of experts with about one trillion parameters, 49 billion active, a one-million-token context and native multimodal input.

When were the weights released?
October 29, 2026, three weeks after an API-only preview that began on October 6.

What does it cost through the API?
$0.70 per million input tokens and $2.10 per million output tokens, according to published listings.

What licence does it use?
The Mistral Ecosystem Open-Weight licence, a modified MIT licence with a clause about offering the model as a hosted service. Read it before building a commercial service around the model.

What hardware do I need?
An eight-GPU class server. NVFP4 weights are roughly 500 GB and FP8 about 1 TB by simple arithmetic, before the context cache.

How good is it compared with closed models?
Mistral says it leads open models on several coding benchmarks. Early Artificial Analysis testing put it well behind the OpenAI and Anthropic flagships.

Verdict

If you need a frontier-scale model you can host yourself under European control, especially for coding or defensive security, Mistral Large 4 is the one to evaluate. The released weights are much stronger than the preview, and the licence is permissive for most uses.

The caveat is the gap to the closed frontier, confirmed by independent testing, and the model-as-a-service clause that matters if you plan to resell access. The hardware requirement is substantial, and many benchmark claims are Mistral’s own.

Next step: test it through Mistral’s API on your real coding or security tasks, read the licence with your legal team, and only then size an eight-GPU deployment.

Sources and Further Reading

On this site: plan serving with LLM inference servers compared, secure agents with prompt injection and agent security, compare coding tools in AI coding assistants in 2026, and see the market in AI models compared.

2 Comments

Comments are closed.