Codestral: The New AI Coding Assistant with Licensing Concerns

Mistral AI has recently released Codestral 22B V0.1, an open-weight generative AI model designed specifically for code generation tasks across more than 80 programming languages, including Python, Java, and JavaScript. Codestral is touted to not only complete functions and write tests but also significantly reduce errors, making it an exciting prospect for developers seeking an AI assistant to help them write better software faster.

From completing functions to writing tests and filling in partial code, Codestral promises to save developers significant time and effort. As usual, the AI community has shared their initial thoughts and discussions in response to this launch.

AI specialists are excited about this new language model, and the initial feedback is positive. Users have found Codestral fast and effective, excelling in various coding tasks. For example, on the chat.mistral.ai platform, the model's speed and accuracy with coding snippets have been highlighted. In another instance, its effectiveness as a coding assistant was particularly noted for its fast performance with a 24GB VRAM setup.

Further positive feedback came from users who tested the model for instruction-based tasks. These tasks included coding a 3D dice roll app, a Snake game, and a visually appealing calculator with animations. Users emphasized that the model performed exceptionally well even with low quantization settings.

However, alongside the excitement, there is significant concern about the licensing under which Codestral was released. The model is licensed under the new Mistral AI Non-Production License, which only restricts its use to research and testing purposes.

Some community members acknowledged the necessity of this restrictive license, given the high costs associated with developing and maintaining such models. Training and maintaining advanced models are costly, and the restrictive license might be necessary to sustain its operations and possibly transition to more permissive licensing in the future. This strategic move is seen by some as a way to balance immediate financial needs with long-term goals. There is hope that Mistral AI might eventually switch to a more permissive license like Apache once a new state-of-the-art model is developed.

Despite understanding the rationale, many users expressed disappointment with the non-commercial restriction. Some felt that the model's utility was significantly limited, making it impractical for real-world applications or self-hosting for commercial projects. The strict license means that the model can only be used for experimentation and hobby projects, which some users find unappealing and not reflective of real-life use cases. This inability to use Codestral's outputs in commercial projects or self-host the model for such purposes has been a major point of contention. Many developers feel that while the model's capabilities are impressive, the restrictive license diminishes its practical value, confining its use to experimental and non-commercial contexts.