Summary

  • French AI start-up Mistral has teamed up with open-source start-up All Hands AI to create Devstral, a small, open-source, licenced 24-million parameter language model for software engineering.
  • While it lacks the capabilities of larger models, such as GPT-4, it has been trained to be an engineering agent, understanding contextual clues across files, navigating large codebases, and solving problems.
  • It achieves a score of 46.8% on the SWE-Bench Verified benchmark, ahead of all other open-source models, and surpasses GPT-4-mini by 20%.
  • As it is only 24 billion parameters it can be run on a laptop and open-source platforms including Hugging Face, Ollama, Kaggle, LM Studio and Unsloth.
  • It is available through an API or locally and has been specifically adapted for use with All Hands AI’s OpenDevin scaffolding for code-focused agents.

By Carl Franzen

Original Article