24B dense (23,572,403,200 parameters)
128K tokens (131,072 configured)
Apache 2.0
Original text-only agentic coding research preview, fine-tuned from the Mistral Small 3.1 family with the vision encoder removed; not the later Devstral Small 2507 or Devstral 2. A 40-layer dense decoder-only Transformer with grouped-query attention (32 query heads, 8 KV heads) and a 131,072-token Tekken vocabulary. The joint technical report describes supervised OpenHands CodeAct trajectories generated in SWE-Gym, unit-test-based patch checks, two stages of increasingly strict trajectory filtering, then additional model rollouts and policy optimization; selected natural-language data retain general language understanding. Designed for tool-driven repository exploration, multi-file edits and real GitHub issue resolution. Publishers report 46.8% SWE-bench Verified resolution with OpenHands on 500 issues; the report specifies bash/file-edit tools, disabled web browsing and an iterative protocol allowing up to three attempts, not standalone no-tool code generation or single-attempt pass@1. Larger open-model comparisons use the same scaffold/protocol; closed-model comparisons use different scaffolds. The first-attempt ablation scores 43.0%, and reducing the iteration limit from 50 to 30 lowers resolution to 36.8%; increasing it to 100 does not improve 46.8%. Official BF16 and GGUF packages support local deployment; the single-RTX-4090/32GB-Mac claim depends on reduced-precision packaging, since BF16 parameter data alone occupy approximately 47.1GB. No image input after vision removal. Release-date review remains open: Mistral explicitly offers downloads starting May 21, while the collaborator page is dated May 20; exact first public availability is unresolved.