Mistral AI has officially launched its second generation of open-source coding model family: Devstral2 and Devstral Small2. The core highlight of this release is the breakthrough performance of the flagship model, while the company has adopted a differentiated licensing strategy for models of different scales.
Flagship Model Devstral2: A Leap in Performance and Cost-effectiveness
The flagship model Devstral2 has an impressive 123 billion parameters. In the key SWE-Bench Verified benchmark test, it achieved a score of 72.2%, surpassing the performance of many similar open-source models.
Mistral AI emphasizes that Devstral2 excels in cost-effectiveness, claiming it is seven times more cost-effective than Claude Sonnet. However, running this model requires high-performance hardware, specifically four H100-level GPUs.
At the same time, the smaller parameter model Devstral Small2 has 24 billion parameters and is designed for local optimization on standard hardware. This lightweight model highlights its efficiency through its performance-to-size ratio.
Licensing Restrictions: Commercial Threshold for High-Income Companies
Mistral AI has applied different licensing terms to the two models:
Devstral Small2: Uses a free Apache 2.0 license, which is highly open.
Devstral2: Uses a "modified MIT license", which sets a revenue cap. Companies with monthly revenue exceeding 20 million dollars, whether using the original Devstral2 or any fine-tuned or derivative version, must purchase a commercial license or access the model through a API.
All models are available on the Hugging Face platform.
Launch of Vibe Command-line Tool to Simplify Workflows
In addition to the model updates, Mistral AI has also launched an open-source command-line tool Mistral Vibe. This tool allows developers to automatically modify code and fix errors across multiple files using natural language, aiming to automate workflows.
Devstral2 is currently available for free via API, with a future pricing of 0.40 USD per million input tokens
