$ cat wiki/models/devstral-2.md
Devstral 2
modelupdated 2026-07-21created 2026-05-17
Spec
| Attribute | Value |
|---|---|
| Developer | Mistral AI |
| Released | 2026-05 (exact day not stated in source) |
| Announced | 2026-05 |
| Context window | 256K (both variants) |
| Pricing | Devstral 2: $0.40 in / $2.00 out per M tokens; Small 2: $0.10 / $0.30 (post-free period) |
| License | Devstral 2: Modified MIT; Small 2: Apache 2.0 |
| Availability | API; open weights |
Variants
| Field | Devstral 2 | Devstral Small 2 |
|---|---|---|
| Parameters | 123B | 24B |
| SWE-bench Verified | 72.2% | 68.0% |
| Deployment | ≥4 H100-class GPUs | DGX Spark / GeForce RTX |
| Multimodal | Image inputs ✓ | Image inputs ✓ |
Release Date
May 2026 (exact date TBD — flagged in sources as a May 2026 release)
Key Claims
- 7× more cost-efficient than Claude Sonnet on real-world tasks (Mistral's claim)
- Devstral Small 2 runs on consumer hardware (RTX-class GPUs, DGX Spark) — local deployment is viable
- Both models are multimodal (image inputs) — can power visual/multimodal agents
Compared To
| Model | SWE-bench Verified | Open-weight |
|---|---|---|
| Devstral 2 (123B) | 72.2% | Yes (modified MIT) |
| Devstral Small 2 (24B) | 68.0% | Yes (Apache 2.0) |
| Claude Sonnet (API) | ~70%+ (est.) | No |
| Grok Build | Not reported | No |
Use Cases
- Agentic software engineering (SWE-bench tasks)
- Codebase exploration and modification
- Powers devstral-2#Mistral-Vibe-CLI
Mistral Vibe CLI
Open-source CLI coding agent built on Devstral:
- License: Apache 2.0
- GitHub: github.com/mistralai/mistral-vibe
- Natural language → file manipulation, code search, version control, shell commands
- Project-aware context (Git status, file structure auto-scan)
- Integrates with IDEs via the Agent Communication Protocol (ACP)
- Available as an extension for the Zed editor
- Partnered with Kilo Code and Cline
Mistral Vibe CLI competes directly with Grok Build (xAI) and Claude Code (Anthropic).
Related
- Mistral AI
- Mistral Medium 3.5 — general-purpose model, same release cycle
- Grok Build — competing agentic coding CLI (xAI)
- Agents (LLM Agents) — agentic coding is the primary use case