Mistral Large 4 Goes Open-Weight at 1 Trillion Parameters
Mistral Large 4, a 1-trillion-parameter open-weight model, is now available with $0.70 per million input tokens. Here is what it means for self-hosting.
PARIS (BYTEVYTE) - Mistral has released the open weights of Mistral Large 4, a 1-trillion-parameter multimodal model, giving enterprises and governments a frontier-scale option to run on their own hardware. The release follows a preview about three weeks earlier and sits alongside hosted API pricing of $0.70 per million input tokens and $2.10 per million output tokens. For a European lab that has raised about $6 billion in three years, the move tests whether sovereign deployment can hold up against rivals committing hundreds of billions of dollars to compute.
The model uses a hybrid instruct-and-reasoning Mixture-of-Experts design, with 52 billion of its 1 trillion parameters active at a time. It is natively multimodal and handles more than 160 languages. Users choose among three reasoning modes, Low, Medium and High (the default), to trade cost against performance. Mistral trained the model on 3,800 NVIDIA Grace Blackwell GPUs in European infrastructure.
Weights are available through Mistral Studio, the Mistral API, OpenRouter, Hugging Face, Microsoft Foundry and NVIDIA NIM. A quantised build runs on four NVIDIA B200 or B300 GPUs, which is the figure that determines the hardware entry cost for self-hosting. Mistral also offers a cybersecurity variant for penetration testing and red-teaming to vetted partners, including Orange Cyberdefense.
Mistral Large 4 against closed and Chinese rivals
Mistral positions the model as a Western alternative to large Chinese open models such as GLM-5.3 and Kimi K3. The company reports that Mistral Large 4 matches top competitors on coding-agent benchmarks while using roughly 47% fewer tokens. Other vendor-reported results include 19.7% on FrontierCode 2.0, 51% on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests. These figures are company-reported, and the sources reviewed contain no independent reproduction of them.
The sources do not compare the API rates with named closed-model prices, so the cost gap cannot be quantified from this material. The asymmetry in capital is clear, however. Mistral has raised roughly $6 billion over three years, while closed frontier labs are spending hundreds of billions on compute. A lab with that budget cannot match capacity, so its leverage rests on price per token and on deployment terms that closed APIs do not offer.
Self-hosting versus closed APIs
Enterprise buyers face a choice between paying per token to a hosted API and running weights on their own infrastructure. Self-hosting keeps data inside the buyer's perimeter, which is the sovereignty argument at the core of the release. It also moves the costs onto the buyer: at least four B200 or B300-class GPUs for the quantised build, plus the power and cooling those systems require. The sources give neither a hardware price nor a power draw for that setup, so total cost of ownership cannot be set against API spend here.
Those power and cooling loads deserve more weight than a line item. Many AI data centers run on-site gas turbines that are loud for nearby residents, gas-fired power carries high CO2 emissions, and cooling these facilities uses large volumes of water. A private deployment brings a share of that footprint onto the buyer's own site, so the environmental and neighbourhood cost belongs in the self-hosting calculation.
Licensing adds a second constraint. The model ships under the Mistral Ecosystem Open-Weight License, a modified MIT license that requires separate commercial terms for Model-as-a-Service providers. Some developers have debated whether that modified license qualifies as fully open, and the sources do not resolve the question.
Enterprise teams should benchmark Mistral Large 4 on their own workloads before choosing between the API and self-hosting, because the performance figures in the sources are vendor-reported. Any team pricing a self-hosted deployment should include the four-GPU footprint, its power and cooling, and the local impact on the site next to the per-token API rate. The next check is whether independent evaluations confirm the 47% token reduction.
While we strive for accuracy and rigor, ByteVyte can make mistakes. Readers are advised to verify all information independently and conduct due diligence before making decisions based on this reporting. We accept no liability for errors, omissions, or decisions made in reliance on this content.
Sources
Mistral Large 4 (Le Chonk) is now open weight
Related Articles
- Mistral Large 4 Preview Tests Europe's Sovereignty Bet
- DeepSeek V4.1-Flash Retires V4 Pro and Collapses Its Model Tiers
- Kimi K3 on Amazon Bedrock Brings a 2.8T Open Model to AWS
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.