Developers experimenting with model-backed agents that inspect codebases and edit multiple files.
The model is documented as a software-engineering fine-tune of Mistral Small 3.1, using a text-only agent setup.
The vision encoder was removed, so this model does not inherit Image Understanding from its base. It needs an agent scaffold and execution tools; weights alone do not run code. Historical SWE-bench scores and rankings are not used as a current best-model claim. Hardware examples do not establish universal memory minima.
Local package without a hosted subscription or per-token inference charge. Hardware, electricity and setup remain separate. License: Apache 2.0.
Install a compatible Ollama version, review the model license, then run: ollama run devstral:24b. API access means the local runtime endpoint, not an included hosted API.
Tags
© Chagto 2026. All rights reserved