The latest model from Mistral AI — Codestral 25.01 — promises even greater speed and capabilities, setting a new standard for coding models in its class.
The latest model from Mistral AI — Codestral 25.01 — promises even greater speed and capabilities, setting a new standard for coding models in its class.
With its improved architecture and updated tokenizer, Codestral 25.01 delivers twice the speed of code generation and completion than its predecessor. When it comes to solving tasks such as fill-in-the-middle (FIM), Codestral 25.01 is now the leader and best-in-class SOTA model, writes Developer Tech News.
To understand how powerful Codestral 25.01 is, Mistral AI compared its performance to other coding models with less than 100 billion parameters, which are considered to be the best at FIM tasks.
In HumanEval tests focused on Python : Codestral 25.01 scored 86.6%, beating its 22 billion predecessor Codestral 2405 and even competitors like Codellama 70B Instruct and DeepSeek Coder 33B Instruct.
For JavaScript : Codestral 25.01 achieved an outstanding HumanEvalFIM score of 87.96%, maintaining a clear lead over the competition.
Across all FIM pass@1 tasks: Codestral 25.01 received an average score of 95.3%, outperforming industry benchmarks and APIs such as OpenAI's GPT-3.5 Turbo.
Codestral 25.01 is available through leading APIs including Vertex AI on Google Cloud, as well as in closed preview on Azure AI Foundry, and will soon be available on Amazon Bedrock. The flexibility of the solution ensures that developers and enterprises can integrate Codestral 25.01 into their workflows.
Additionally, Mistral offers open access to the Codestral documentation and API on its la Plateforme portal. Developers can experiment with integrations, ensuring that the model’s capabilities match their specific needs.
For those interested in trying out Codestral 25.01, the model is available through IDE plugins for platforms such as VS Code and JetBrains.
