Mistral AI has released Codestral Mamba, an open-source language model designed for efficient code generation and reasoning, offering linear time inference and the ability to handle infinitely long sequences.
This release follows the company's Mixtral family of models and offers several key advantages over traditional Transformer models. It provides linear time inference, allowing for quick responses regardless of input length, and has the theoretical ability to model sequences of infinite length. The model also demonstrates advanced code and reasoning capabilities, performing on par with state-of-the-art transformer-based models.
The model has been rigorously tested for in-context retrieval capabilities up to 256,000 tokens, making it an ideal local code assistant. Mistral AI has made Codestral Mamba available under the Apache 2.0 licence, encouraging free use, modification, and distribution within the AI community.
For easy testing, Mistral AI has made Codestral Mamba accessible on their platform, la Plateforme, alongside its larger counterpart, Codestral 22B. While Codestral Mamba is open-source, Codestral 22B is available under commercial and community licences for different use cases.
With the release of Codestral Mamba, Mistral AI continues to push the boundaries of AI model architecture. This innovative approach to language modelling, particularly for code-related tasks, could significantly impact the field of AI-assisted programming and open new avenues for research in AI architectures.