Mistral Large 4 launches with 1 trillion parameters and 49 billion active weights

Mistral AI released Mistral Large 4, a new open-weight model designed for developers who need high-capacity reasoning without the cost of proprietary systems. The release matters because it offers a 1 trillion parameter architecture that activates only 49 billion parameters, allowing for efficient inference on standard hardware. The sparse architecture allows users to execute complex tasks, such as code generation and data analysis, with significantly lower compute costs compared to dense models. Buyers looking for a European-based alternative to US-centric models now have a viable option that supports over 160 languages and native multimodal inputs.

Sparse MoE model offers European data residency and multimodal support

The model, internally dubbed "Le Chonk," uses a sparse Mixture-of-Experts (MoE) architecture to manage its massive scale. It supports a context window of 524,288 tokens and can output up to 262,144 tokens in a single response. Mistral AI trained the model from scratch using 4,000 NVIDIA Grace Blackwell GPUs over a two-month period in European servers. The use of European servers supports data residency requirements for users needing to comply with regional privacy regulations.

Performance benchmarks show mixed results across different workloads. With a score of 62% on DeepSWE v1.1, the model demonstrates strong capabilities in software engineering tasks. It also achieved a score of 38 on the Artificial Analysis Intelligence Index and 50 on the Cyber Index. However, it trails behind competitors like GLM-5.3 and Kimi K3 on several other metrics. The model does lead on the Harvey's Legal Agent benchmark, which aligns with its European regulatory focus.

Pricing for Mistral Large 4 is set at $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. The standard cost per Intelligence Index task is $1.13. The pricing structure positions the model as a cost-effective option for high-volume API usage. Developers can access the model through standard API endpoints, with availability extending to users in the US and EU regions.

Mistral AI claims the model is the best open-weight option from the US or Europe on aggregated benchmarks, though this assertion relies on their own announcement data. Independent verification of these aggregated scores remains limited in the current reporting. The release represents a notable development for European AI infrastructure by offering a transparent alternative to closed models. Users are advised to test the model against their specific workloads to verify performance claims.