Source-level technical characteristic used for deployment comparison.
Source-verified model profile · 1 October 2026
Devstral Small 2 24B Instruct 2512
Devstral Small 2 24B Instruct 2512 is an open-weight model from Mistral AI focused on software engineering, repository-scale coding and agentic development. The profile separates model facts, license terms, hardware, local/self-hosted operation and EU data-residency questions.
OPEN WEIGHTSAPACHE 2.0EUROPEAN UNIONEU SELF-HOSTING POSSIBLE
Direct answer
What is Devstral Small 2 24B Instruct 2512?
Devstral Small 2 24B Instruct 2512 is a Devstral 2 open-weight model published by Mistral AI. It is aimed at software engineering, repository-scale coding and agentic development, with 24B-class, Up to 262,144 in published generation configuration of context and Text/code → text/code. The recorded license is Apache 2.0.
The exact checkpoint, precision and runtime matter. Quantized or optimized variants can reduce memory substantially, while long context and multimodal inputs increase runtime memory.
Official model source ↗Provider origin🇫🇷 France
RegionEuropean Union
Parameters24B-class
Active computeDense
ContextUp to 262,144 in published generation configuration
ModalitiesText/code → text/code
LicenseApache 2.0
VerificationSource-verified · 2026-10-01
Technical profile
What matters for selection.
Source-level technical characteristic used for deployment comparison.
Source-level technical characteristic used for deployment comparison.
Source-level technical characteristic used for deployment comparison.
Source-level technical characteristic used for deployment comparison.
Hardware & runtime
Can it run locally?
High-memory workstation or GPU server class; quantization can reduce the footprint.
Mistral publishes vLLM/mistral-common integration; use the model-specific chat/system prompts for agentic coding.
Do not size hardware from parameter count alone: precision, context length, KV cache, batch size and modality encoders can change the real memory envelope.
License
Commercial-use check.
The recorded license is Apache 2.0. Open weights describe technical availability, not a blanket statement that every use, derivative, redistribution or hosted service is unrestricted.
For production, retain the exact license and checkpoint revision and review any separate usage policy.
Technical comparison only — not legal advice.
EU deployment lens
Origin, residency and compliance are separate.
🇫🇷 France is the provider origin recorded for this model. It is not a statement about training-data origin or where your operational data is processed.
EU/EEA self-hosting is technically possible when weights, inference, RAG, embeddings, logging, monitoring and backups are kept on the selected infrastructure. Any external services must be assessed separately.
Self-hosting is not itself a GDPR or AI Act compliance badge.
EU deployment & data-residency guide →FAQ
Common questions.
What is Devstral Small 2 24B Instruct 2512?
Devstral Small 2 24B Instruct 2512 is an open-weight model from Mistral AI focused on software engineering, repository-scale coding and agentic development.
Can Devstral Small 2 24B Instruct 2512 run locally?
High-memory workstation or GPU server class; quantization can reduce the footprint.
Can Devstral Small 2 24B Instruct 2512 be deployed in the EU?
Yes, technically, when the downloadable weights and complete inference/data stack are operated on EU/EEA infrastructure.
Can Devstral Small 2 24B Instruct 2512 be used commercially?
The recorded license is Apache 2.0; review the exact model terms before production use.
Primary source
Verify the exact checkpoint.
Publisher documentation remains authoritative for revision-specific files, runtime settings, context behavior and license links.
Mistral AI · official model page ↗