Mistral's code model, CodeStral, is back with an upgrade!
Now it stands shoulder to shoulder with DeepSeek V2.5 and Claude 3.5, tying for the top spot in the Copilot Arena.
The context window has also expanded to 256k—8 times its previous size.
According to the introduction, the new CodeStral (2501) adopts a more efficient architecture and tokenizer, delivering generation speeds approximately twice as fast as the previous version.
Across multiple benchmarks, the 2501 version has achieved state-of-the-art (SOTA) results, with particularly notable performance in code completion (Fill-in-the-Middle, FIM) tasks.
Ty Dunn, co-founder of Mistral’s partner Continue.dev, remarked that CodeStral 2501 represents a significant breakthrough in the FIM domain.
