Codestral Mamba

Mistral AI · 2 years ago

Mistral released Codestral Mamba, a code-focused AI model based on the Mamba architecture rather than Transformers, available for free use and modification. The model contains 7.3 billion parameters, supports context windows up to 256,000 tokens, and performs comparably to state-of-the-art Transformer-based code models while offering linear-time inference. Developers can now deploy it locally or via Mistral's platform for code assistance tasks, with support coming to llama.cpp for easier integration.