123B nominal Devstral 2 (125,025,988,608 learned parameters); 24B Devstral Small 2 (24,011,361,280 including vision and connector); dense
256K (262,144 tokens) supported for both; Small 2 configuration ceiling: 393,216 tokens
Devstral 2: Modified MIT, retaining attribution/permission notices; no rights under this license when company/employer global consolidated monthly revenue for the preceding month exceeds $20 million (or equivalent currency), including derivatives/modifications/combined works; a discretionary commercial license or Mistral hosted service is an alternative. Exactly $20 million does not trigger the stated exceeds condition. Devstral Small 2: Apache 2.0 publisher grant; standalone LICENSE file absent from the checked repository.
Two instruction-tuned, FP8 open-weight coding-agent checkpoints released together for tool-driven codebase exploration, multi-file edits, debugging and software-engineering workflows. Devstral 2 is text-only: an 88-layer dense decoder with 96 query heads, 8 KV heads and untied embeddings. Devstral Small 2 adds image inputs through a 24-layer Pixtral vision encoder and patch-merge connector to a 40-layer dense decoder with 32 query heads and 8 KV heads; its model card identifies Mistral-Small-3.1-24B-Base-2503 as the base and describes Ministral-3-style attention-temperature/RoPE scaling. Both advertise 256K context; the Small checkpoint's larger positional ceiling is not a verified 384K supported window. Publisher-reported SWE-bench Verified scores are 72.2% and 68.0%, SWE-bench Multilingual 61.3% and 55.7%, and Terminal-Bench 2 32.6% and 22.5%, respectively. These are agent-system results, not demonstrated bare-model pass@1: the launch sources do not specify the benchmark scaffold, tool-turn budget, sampling, retries or aggregation. In separate human evaluations scaffolded through Cline, Devstral 2 recorded a 42.8% win rate and 28.6% loss rate against DeepSeek V3.2, while Claude Sonnet 4.5 remained significantly preferred. Mistral claims up to 7x better real-world cost efficiency than Claude Sonnet. Released alongside the Apache-2.0 Mistral Vibe CLI, with file, shell, code-search and version-control tools; integrations include Cline, Kilo Code and Zed. Mistral recommends at least four H100-class GPUs for Devstral 2 and describes Small 2 as locally deployable on consumer hardware; single-GPU/32GB-Mac claims depend on quantization, runtime and context allocation. Both are documented as instruction fine-tunes, and a publisher response confirms Small 2 was trained natively in FP8; the complete release-specific data mixture, training budget and optimization stages are undisclosed. No predecessor training recipe is assumed.