What it is
Devstral 2 is a 123-billion-parameter dense coding model from Mistral AI, released on December 9, 2025 together with the Mistral Vibe command-line coding agent. It reads and writes text, supports tool calling and has a 256K-token context. The Hugging Face weights are almost entirely FP8, and Mistral says it needs at least four H100-class GPUs.
The weights use a modified MIT licence. Companies whose monthly revenue, including their group, exceeded $20 million in the previous month get no rights under it, including for fine-tunes and other derivatives, and must use Mistral's hosted service or buy a commercial licence. Everyone must keep a Mistral attribution notice.
Mistral's API charged $0.40 per million input tokens and $2.00 per million output tokens after a free launch period. Mistral deprecated it on May 22, 2026 and retired it from the API on July 31, 2026, naming Mistral Medium 3.5 as the replacement. The weights remain on Hugging Face.
Our review
Devstral 2 is a retired model, and that settles most decisions about it. At launch Mistral reported 72.2% on SWE-bench Verified, 61.3% on SWE-bench Multilingual and 32.6% on Terminal Bench 2, all self-reported; Claude Sonnet 4.5 scored 77.2% and 42.8% in the same table. In a human evaluation run through Cline, Devstral 2 beat DeepSeek V3.2 42.8% of the time and lost 28.6%, but Mistral admitted Claude Sonnet 4.5 was still clearly preferred.
Mistral deprecated it on May 22, 2026 and switched it off on the API on July 31, naming Mistral Medium 3.5 as the replacement. Mistral Small 4, from March 2026, combines instruct, reasoning and coding in one Apache 2.0 model. With the vendor pointing everyone elsewhere, a new project on Devstral 2 starts with a migration already owed.
The licence is the other problem, and it does not go away by staying with Mistral. The $20 million monthly revenue cap covers fine-tunes and third-party derivatives, and Mistral Medium 3.5 carries the same modified MIT terms. Only the smaller models, Devstral Small 2 and Mistral Small 4, are Apache 2.0.
Serving it is expensive for what it now offers. The weights come to about 125 GB, Mistral's floor is four H100-class GPUs, and the model card's own vLLM example spreads it across eight. OpenRouter's catalogue still lists a Mistral-served endpoint at the old $0.40/$2.00 price, which sits oddly with Mistral's retirement table; anyone relying on it is one cleanup away from an outage.
Where it earns its place
A platform engineer at a firm under the revenue cap, already serving Devstral 2 on a 4xH100 node
Keep an in-house coding agent running while a replacement is evaluated
The weights keep working under vLLM with Mistral's tool-call parser, and the only cost is the node. Running the team's own tasks against Mistral Medium 3.5 or Small 4 tells you whether the next refresh can use smaller hardware.
A legal or procurement lead at a company above $20M monthly revenue
Check whether engineers may self-host a model they found on Hugging Face
They may not without a commercial licence from Mistral: the cap uses group revenue for the previous month and covers derivatives. Devstral Small 2 and Mistral Small 4 are Apache 2.0 and avoid the question.
A researcher benchmarking open coding models
Reproduce a late-2025 open-weights baseline on SWE-style tasks
The FP8 checkpoint is a fixed reference point with a published 72.2% SWE-bench Verified score. Budget eight GPUs if following the model card's vLLM example.
Before you commit
- Setup
- vllm serve with Mistral's tool-call parser; the model card's example uses tensor parallelism across 8 GPUs, and Mistral recommends a temperature of 0.2.
- Integrations
- OpenAI-compatible through vLLM; Mistral Vibe CLI, Cline and Kilo Code were launch partners, and Vibe can point at a local server through its config.toml.
- Data and privacy
- Modified MIT: no rights for companies above $20M global monthly revenue in the previous month, including for derivatives, and an attribution notice is required. Self-hosted, code stays on your hardware.
- Cost at scale
- Self-hosting needs about 125 GB for weights, so a four-GPU 80 GB node is the floor before KV cache. On Mistral's API, 1B input and 100M output tokens a month would have cost about $600 at list price.
- Limits
- Retired from Mistral's API on July 31, 2026. Text only, 256K-token context; 32.6% on Terminal Bench 2 against Claude Sonnet 4.5's 42.8% in Mistral's own table.
What's good
- Open weights still available, in FP8
- 72.2% SWE-bench Verified at launch (self-reported)
- 256K-token context with tool calling
What's not
- Retired from Mistral's API on July 31, 2026
- No rights for companies over $20M monthly revenue, including derivatives
- Dense 123B model needs four to eight data-centre GPUs
- Trailed Claude Sonnet 4.5 in Mistral's own tests
Best for
Existing self-hosted users who are not ready to migrate yet.
Alternatives
Devstral 2: common questions
- Is Devstral 2 worth it?
- A solid coding model for late 2025 that its own maker has retired; new projects belong on Mistral Medium 3.5 or a newer open model. We rate it 3.1 out of 5.
- How much does Devstral 2 cost?
- Open weights under a modified MIT licence (not for companies over $20M monthly revenue). Retired from Mistral's API on July 31, 2026.
- What are the downsides of Devstral 2?
- Retired from Mistral's API on July 31, 2026; no rights for companies over $20M monthly revenue, including derivatives; dense 123B model needs four to eight data-centre GPUs; trailed Claude Sonnet 4.5 in Mistral's own tests.
- Who is Devstral 2 best for?
- Existing self-hosted users who are not ready to migrate yet.
- How much does Devstral 2 cost for a team or in production?
- Self-hosting needs about 125 GB for weights, so a four-GPU 80 GB node is the floor before KV cache. On Mistral's API, 1B input and 100M output tokens a month would have cost about $600 at list price.
- Is Devstral 2 safe for company data?
- Modified MIT: no rights for companies above $20M global monthly revenue in the previous month, including for derivatives, and an attribution notice is required. Self-hosted, code stays on your hardware.
- What are the best Devstral 2 alternatives?
- The closest alternatives we have reviewed are Mistral Medium 3.5, Mistral Small 4, GLM-5.3-Flash and Devstral Small 2.
Answers checked Oct 22, 2026.