Mistral releases open weights for its AI model

Mistral, a French startup, made the open weights for its artificial intelligence model Mistral Large 4, also known as Le Chonk, available on October 29, as previously announced. These weights, which are the adjustments learned during training, are now accessible on the Hugging Face website, a platform for sharing machine learning models.

Model uses a Mixture of Experts architecture

The Mistral Large 4 model uses a Mixture of Experts (MoE) architecture, which includes 128 "experts." For each token, only four of these experts (plus one common expert) are used. The model has 1,000 billion parameters in total, but only 49 billion are used per token, with 52 billion if the input and output layers are included. This is an improvement over its predecessor, Mistral Large 3, which had 675 billion parameters.

Mistral refined the model with reinforcement learning

Between the announcement of Mistral Large 4 and the release of its weights, Mistral continued to refine the model using reinforcement learning, focusing on areas like cybersecurity and code. The company reported "considerable improvements" and said progress "continues to accelerate, without any sign of saturation."

Model shows improved performance in benchmarks

According to Mistral's figures, the model's score on Terminal-Bench 4.0, a command-line programming test, increased from 28% to 41%, just one point ahead of the Chinese GLM-5.3 (40%). On the code index from Artificial Analysis, it achieved 48%, matching the GLM-5.3, while using 47% fewer tokens per task. It outperformed several other models on Terminal-Bench but still trails behind the best open Chinese models on the same firm's overall index.

Ecosia switches to Chinese models

Ecosia, a search engine, began replacing Mistral's models with Chinese alternatives such as Qwen, Kimi, and GLM at the start of October. Christian Kroll, Ecosia's head, criticized Mistral for the quality of its models and a delay of about a year compared to its competitors.

Mistral lowers API prices but tightens licensing

Mistral reduced the API price for its model, with input costs now at 0.70 dollars (approximately 0.63 euros) per million tokens and output costs at 2.10 dollars (approximately 1.89 euros). This is a reduction from the standard previous price of 1.36 dollars (approximately 1.22 euros) and 4.18 dollars (approximately 3.75 euros), respectively.

New license requires additional agreements for commercial use

The license for Mistral Large 4, called the Mistral Ecosystem Open-Weight License, is less permissive than the Apache 2.0 license used for Mistral Large 3, which allowed all uses, including commercial ones. The new license allows commercial use but requires companies that resell access to the model as a service to sign a separate license with Mistral. Even the compressed version of Le Chonk requires four Nvidia B200 or B300 data center graphics cards, and the full version requires eight.