self-hosted/ai
§01·model · /models

Devstral Small 2 (24B)

llmactiveApache-2.0

Mistral AI's dedicated agentic-coding model and the first Mistral-family model in the catalog. Devstral Small 2 is a dense 24B model (fine-tuned from Mistral-Small-3.1-24B-Base) purpose-built for software-engineering agents: scoring 68.0% on SWE-bench Verified and 55.7% on SWE-bench Multilingual, it matches much larger models like GLM-4.6 (355B) while running locally on a single 24 GB GPU or a 32 GB Mac. Apache-2.0 licensed, 256k context. Designed to be driven by agentic coding harnesses (Mistral Vibe, OpenHands, Cline, SWE-agent, Claude Code) over an OpenAI-compatible API. The released checkpoint also carries a vision encoder for image-aware coding tasks.

§02·GPUs that run this model
8 total
GPUVRAMSeriesBest speedMin VRAMWorksBenchmarksRecipe
Apple M2 Max64GBapple~0recipecheck ↗
Apple M3 Max48GBapple~0recipecheck ↗
RTX 309024GB30~0recipecheck ↗
RTX 3090 Ti24GB30~0recipecheck ↗
RTX 408016GB40~0recipecheck ↗
RTX 409024GB40~0recipecheck ↗
RTX 509032GB50~0recipecheck ↗
RX 7900 XTX24GBamd~0recipecheck ↗

benchmarked·~ runs via recipe (not benchmarked)· untested·doesn't fit