Mistral AI is Europe's leading AI lab, producing a comprehensive family of open-weight models anchored by Mistral Medium 3.5 (128B dense, 77.6% SWE-bench Verified, 256K context) and Mistral Large 3 (675B MoE, #2 open-source non-reasoning model on LMArena) -- all under the Apache 2.0 license, with a strong EU data sovereignty story.
Why It's in Trial
Mistral's December 2025 and April 2026 releases elevated the company from "interesting European alternative" to genuine contender:
- Mistral Medium 3.5 (128B dense): 77.6% on SWE-bench Verified (April 29, 2026) — a unified 128B model that consolidates Devstral 2, Magistral, and Devstral Small 2 into a single open-weight release. Now the default model in both Vibe and Le Chat.
- Devstral 2 (123B): 72.2% on SWE-bench Verified -- one of the top open-weight scores for code agents (superseded by Medium 3.5 but still widely used)
- Devstral Small 2 (24B): 68.0% on SWE-bench Verified while running on consumer hardware -- a remarkable efficiency story
- Mistral Large 3 (675B MoE, 41B active): #2 among open-source non-reasoning models on LMArena, trained on 3,000 NVIDIA H200 GPUs
- Apache 2.0 license: The most permissive licensing in the space -- commercial use, modification, redistribution all unrestricted
- Mistral Vibe CLI: Native ACP (Agent Client Protocol) support, making Mistral one of the few model providers with their own coding agent in the ACP registry
- Broad ecosystem presence: Referenced across this radar -- Ollama, Groq, Vercel AI SDK, Hugging Face, Azure AI
The Model Family
| Model | Parameters | Release | Strength |
|---|---|---|---|
| Mistral Medium 3.5 | 128B dense | Apr 2026 | Unified flagship: 77.6% SWE-bench, 256K context, multimodal, configurable reasoning |
| Mistral Large 3 | 675B MoE (41B active) | Dec 2025 | Flagship general-purpose, 256K context |
| Devstral 2 | 123B | Dec 2025 | Code agents, 72.2% SWE-bench (superseded by Medium 3.5) |
| Devstral Small 2 | 24B | Dec 2025 | Local code agents, 68.0% SWE-bench |
| Ministral 3 (14B/8B/3B) | 3B-14B | Dec 2025 | Dense models with reasoning variants |
| Codestral 25.08 | — | Aug 2025 | FIM-optimized for IDE integration |
| Mistral Medium 3 | — | May 2025 | Mid-tier general-purpose (superseded by Medium 3.5) |
European Data Sovereignty
For teams subject to GDPR, EU AI Act, or European data localisation requirements, Mistral offers a genuinely differentiated option:
- Headquartered in Paris, France -- EU-based company subject to European regulation
- La Plateforme: Mistral's own inference platform for European-hosted API access
- Open weights: Self-host on your own EU-based infrastructure for full data control
- Valued at $14B+ as of late 2025, with significant European institutional backing
Mistral Vibe CLI
Mistral released Vibe, a command-line interface for AI-assisted software development with native ACP (Agent Client Protocol) support. This makes Mistral one of the few model providers building their own agentic coding tool -- see the Agent Client Protocol (ACP) entry for context on the broader protocol ecosystem.
Key Characteristics
| Property | Value |
|---|---|
| Flagship model | Mistral Medium 3.5 (128B dense, Apr 2026) |
| Best coding model | Mistral Medium 3.5 -- 77.6% SWE-bench Verified |
| Context window | Up to 256,000 tokens |
| License | Apache 2.0 |
| Provider | Mistral AI (Paris, France) |
| Latest releases | April 2026 (Medium 3.5) |
| API | La Plateforme |
| Weights | mistralai/Mistral-Medium-3.5-128B |