Published: 29 October 2026
Mistral AI published the weights of Mistral Large 4 on Hugging Face on 29 October, three weeks after it launched the model as a preview. Any organisation may now download the 1 trillion parameter model, adapt it and use it commercially. Only companies that sell it as a model service need a separate licence from Mistral. The catch is hardware: even a compressed version needs a server with four Nvidia B200 or B300 GPUs, so most teams will start through the API.
Only companies that sell Large 4 as a service need a commercial licence
Large 4 comes under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. The model is free to use, modify, fine-tune and commercialise, with one condition. Businesses that run a model-as-a-service offering need a separate commercial licence before they use it commercially.
For a bank, ministry or manufacturer that runs the model for its own staff, there is no licence fee. A Dutch cloud or hosting company that wants to offer Large 4 to its customers needs a contract with Mistral first. The announcement does not say where a managed setup for one client ends and a model service begins. If your plans sit in that grey zone, read the licence file on the Mistral Large 4 page on Hugging Face before you build on it.
Running Large 4 in-house takes a GPU server, not a workstation
Large 4 is a mixture-of-experts model. It has 1 trillion parameters in total, but uses only about 52 billion of them for each token (a word or part of a word). That makes it faster than its size suggests, yet all the weights still have to fit in GPU memory.
The Next Web reports that a quantised version, compressed to use less memory, runs on four Nvidia B200 or B300 GPUs. Mistral publishes such a version for Nvidia’s Blackwell chips, and its own deployment examples on Hugging Face use eight GPUs. The model reads up to 1 million tokens at once and accepts both text and images.
Smaller versions are on the way. Mistral told The Next Web it will release distilled models “in the coming months”, including one that runs on a laptop. It gave no dates.
Large 4 ties GLM-5.3 on coding agents, but Chinese open models still score higher overall
Mistral spent the three weeks since the preview on extra reinforcement learning, a training method that rewards good answers. On Terminal-Bench 4.0, a test of coding agents that work in a terminal, Large 4 went from 28% at the preview to 41%. Mistral’s chart puts GLM-5.3 at 40% and DeepSeek V4.1 Flash at 31%, The Next Web reported.
On the Artificial Analysis Coding Agent Index, Mistral says Large 4 scores 48% and is effectively tied with GLM-5.3. It gets there with 47% fewer tokens per task, which lowers the bill per task and not just per token. On the broader Artificial Analysis Intelligence Index, Large 4 scores 26 according to Mistral’s model card. That is the highest score of any open-weight model from a Western lab, but Chinese open-weight models score higher.
Artificial Analysis runs its tests independently, but Mistral chose which results to show. Treat these numbers as a reason to run your own test, not as proof that Large 4 fits your work.
Does the Mistral API keep Large 4 traffic in the EU?
Only when you use Mistral’s EU endpoint, and that costs 10% more. Mistral’s default API address makes no promise about where your prompts are processed. The EU endpoint sends processing to European servers at 1.1 times the list price. Mistral’s documentation says not every model runs in every region, and it does not list which ones do. Ask the EU endpoint for its model list before you rely on it. Account details, billing and usage data can still be handled outside the region.
Large 4 is available today through these routes:
- Mistral API: $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens
- Model ID in the API: mistral-large-4
- Also on Mistral Studio, Vibe Code, OpenRouter, Nvidia’s hosted endpoints and Microsoft Foundry
- Weights on Hugging Face, with ready setups for vLLM and SGLang and community builds for llama.cpp
Mistral says it trained Large 4 from scratch in Europe. The Next Web reports that the run used 3,800 Nvidia Grace Blackwell GPUs, with the reinforcement learning on Mistral’s own computers in France. The model card lists more than 160 languages, including every official EU language, so Dutch is covered.
ASML and ABN AMRO are among the companies quoted in the announcement. Arnaud Hubaux, who leads ASML’s AI programme, says: “Our early experience with Mistral Large 4 is promising, particularly for use cases where higher model performance matters.”
Mistral is still adjusting the safeguards, and downloaded weights stay as they are
Mistral also gives vetted security firms, such as Orange Cyberdefense, early access to a separate Large 4 variant for penetration tests and security audits.
For the open model, Mistral writes that Large 4 refused attacks on live targets in its tests, but that it could still help with real attacks. “We are modifying our guardrails accordingly, and will continue updating them in the months to come,” the company says. Weights you download today will not receive those changes unless you download a new version. If you host the model yourself, the safeguards around it are yours to set: who may use it, what gets logged and which tools it may call.
Measure Large 4 on your own work through the API before anyone orders GPUs
Some organisations must keep data under their own control, such as government bodies, hospitals and financial firms. For them, Large 4 is a European model close to the strongest open models that they may run on their own hardware without a licence fee. That makes it a serious option for sovereign AI. But the licence is the easy part. A self-hosted Large 4 needs GPU servers, serving software, monitoring and people to keep it updated. In local AI projects, those running costs usually weigh more than the price per token.
For most teams, the sensible order is this. Run Large 4 through Mistral’s API on a set of your own tasks, in Dutch and English, on the EU endpoint if it lists the model. Compare it with the model you use now on quality and on cost per finished task. Consider self-hosting when your volume or your data rules justify a GPU server, or when a hosting partner offers Large 4 with Mistral’s commercial licence in place. If you need a model for laptops or a single GPU, wait for the distilled versions.
For more information, visit the official announcement of Mistral Large 4 on the Mistral AI blog.
