What it is
Devstral Small 2 is a 24-billion-parameter coding model from Mistral AI, released on December 9, 2025 alongside Devstral 2 and the Mistral Vibe command-line agent. It reads text and images, writes text, supports tool calling and has a 256K-token context. According to its model card it is light enough for a single RTX 4090 or a Mac with 32GB of RAM.
Mistral has since wound down the whole Devstral line on its API. Devstral Small 2 was deprecated on February 27, 2026 and retired on March 31, 2026; Devstral 2 followed on July 31, 2026. Mistral points users to Mistral Medium 3.5, and its models page lists no current Devstral model. The Apache 2.0 open weights remain on Hugging Face and work with vLLM, llama.cpp, LM Studio and Ollama.
Our review
Devstral Small 2 is a capable local coding model whose maker has moved on. Mistral retired its API version in March 2026 and the larger Devstral 2 in July, and its models page now lists no Devstral at all. Whatever you build on it, you maintain without a vendor roadmap.
The numbers that made its name are Mistral's own: 68.0% on SWE-bench Verified, 55.7% on SWE-bench Multilingual and 22.5% on Terminal Bench 2, level with the 355B GLM 4.6 on Verified in Mistral's table. Claude Sonnet 4.5 scored 77.2% in the same comparison. Artificial Analysis independently puts it at 8 on its Intelligence Index, 17th of 74 open non-reasoning models between 4B and 40B parameters, where the median is 6.
Its real asset is that it fits. The official weights are FP8, about 24GB, and Unsloth's Q4_K_M quantisation is 14.3GB, so a 24GB RTX 4090 holds the model with room for a useful slice of the 256K context. Mistral's suggested successors, Mistral Small 4 at 119B and Medium 3.5 at 128B, do not fit that hardware at all.
The weakness is time. IBM's Granite 4.2 and Google's Gemma 4 are 2026 Apache 2.0 models built for the same single-GPU slot, and both are newer than anything left in the Devstral line. Benchmarks across vendors use different agent scaffolds, so the honest comparison is a week on your own repository.
Where it earns its place
A security-conscious fintech team of eight engineers
Run a coding agent without sending source code to any cloud
Serve the FP8 weights with vLLM on one 48GB workstation GPU and point Cline or OpenHands at it. Hardware is a one-off cost; the risk is that no vendor will patch or update the model.
A freelance developer on a 32GB MacBook Pro
Get an offline coding assistant for travel and client sites with no internet
Pull devstral-small-2 from Ollama, about 15GB, and use it in LM Studio or Cline. Responses are slower than a hosted model and quality drops on large multi-file refactors.
An engineering manager comparing local models for a 30-person team
Choose a baseline for an internal self-hosted coding service
Devstral's 68% SWE-bench Verified is a reasonable reference point, but with the line retired the safer bet for a multi-year service is a family still receiving releases.
Before you commit
- Setup
- ollama pull devstral-small-2 or serve with vLLM using Mistral's tokenizer flags; Mistral recommends a server and client setup and asks users to report local-serving bugs to the framework.
- Integrations
- vLLM, SGLang, Transformers, llama.cpp, LM Studio and Ollama; coding agents including Mistral Vibe, Cline, Kilo Code, Claude Code, OpenHands and SWE-agent.
- Data and privacy
- Apache 2.0 with no usage restrictions; self-hosted, so code never leaves your hardware unless you choose a third-party host.
- Cost at scale
- No first-party API price exists any more; self-hosting means about 24GB of VRAM for the official FP8 weights, 19.3GB at Q6_K or 14.3GB at Q4_K_M, plus memory for the KV cache at long contexts.
- Limits
- Mistral's API endpoints for every Devstral model are retired; Artificial Analysis rates it somewhat verbose, using 18M tokens on its index against a 16M median.
What's good
- Apache 2.0 licence
- Q4 quantisation at 14.3GB fits one RTX 4090
- 256K-token context and image input
- Works with Cline, OpenHands, Claude Code and Mistral Vibe
What's not
- Mistral's API version retired in March 2026, and no Devstral remains on its API
- No public hosted API listed by Artificial Analysis
- SWE-bench Verified 9 points behind Claude Sonnet 4.5 in Mistral's own table
- No further updates expected
Best for
Developers who want a self-hosted coding agent on one 24GB GPU or a 32GB Mac and are content to maintain a model its maker no longer develops.
Alternatives
Devstral Small 2: common questions
- Is Devstral Small 2 worth it?
- Devstral Small 2 still runs well on one gaming GPU, but Mistral has abandoned the line and nobody now sells it as a service. We rate it 3.4 out of 5.
- How much does Devstral Small 2 cost?
- Free weights under Apache 2.0. Mistral's API version was retired on March 31, 2026, and Artificial Analysis lists no public API.
- What are the downsides of Devstral Small 2?
- Mistral's API version retired in March 2026, and no Devstral remains on its API; no public hosted API listed by Artificial Analysis; sWE-bench Verified 9 points behind Claude Sonnet 4.5 in Mistral's own table; no further updates expected.
- Who is Devstral Small 2 best for?
- Developers who want a self-hosted coding agent on one 24GB GPU or a 32GB Mac and are content to maintain a model its maker no longer develops.
- How much does Devstral Small 2 cost for a team or in production?
- No first-party API price exists any more; self-hosting means about 24GB of VRAM for the official FP8 weights, 19.3GB at Q6_K or 14.3GB at Q4_K_M, plus memory for the KV cache at long contexts.
- Is Devstral Small 2 safe for company data?
- Apache 2.0 with no usage restrictions; self-hosted, so code never leaves your hardware unless you choose a third-party host.
- What are the best Devstral Small 2 alternatives?
- The closest alternatives we have reviewed are Gemma 4, Granite 4.2, Mistral Small 4 and Devstral 2.
Answers checked Oct 22, 2026.