Devstral 2 Guide: Mistral's 123B Dense Coding Agent

Devstral 2 is easy to misread from its headline. "Open-weight 123B coding model from Mistral" suggests a permissively licensed model anyone can build on. The weights are open, but the licence has a revenue clause that rules out a lot of companies, and the smaller sibling released the same day does not have it.

That licence split, more than any benchmark, is what decides which Devstral you can use. This guide covers what the model is, where its numbers sit, and how to choose between the two sizes.

What Mistral shipped in December

Mistral released Devstral 2 on 9 December 2025 alongside Devstral Small 2 and an open-source command-line agent called Mistral Vibe. The two models:

• Devstral 2 — 123B parameters, dense transformer, 256K context
• Devstral Small 2 — 24B parameters, 256K context, accepts image input

Both are aimed at agentic coding: reading a codebase, editing several files, running commands and iterating. That makes them a different tool from Mistral's completion model, covered in the Codestral guide, which fills in code at the cursor rather than running a loop.

Mistral priced Devstral 2 at $0.40 per million input tokens and $2.00 per million output after an initial free period, and Small 2 at $0.10 and $0.30. OpenRouter listed Devstral 2 at the same $0.40 and $2.00 at the time of writing.

Dense, not mixture-of-experts

Most open models at the top of the 2026 coding charts are sparse mixture-of-experts designs with a very large total count and a small active one. Devstral 2 goes the other way: all 123B parameters run on every token.

Mistral's framing is that this makes it much smaller than its competitors, about five times smaller than DeepSeek V3.2 and eight times smaller than Kimi K2 by total parameters. That is true for memory. It is not true for compute per token, where a dense 123B model does more work than a sparse model with 30B or 40B active.

The practical consequences cut both ways. A dense model is simpler to serve, with no expert routing and more predictable memory use, and it quantises in a well-understood way. But Mistral states a minimum of four H100-class GPUs to run Devstral 2, and throughput per GPU will trail a sparse model of similar quality. Dense vs mixture-of-experts covers which shape is easier to run on the hardware you have.

Read the licence before anything else

Devstral 2 is released under what Mistral calls a modified MIT licence. The modification, as quoted by Simon Willison from the licence text, is that you may not exercise any rights under it if the global consolidated monthly revenue of your company, or of your employer, exceeds $20 million for the preceding month. Willison notes that the restriction extends to derivatives and modifications, including third-party ones.

Three things follow from that:

• A company above the threshold cannot self-host the weights under this licence, even internally, without a separate arrangement with Mistral.
• A fine-tune of Devstral 2 inherits the restriction, so a startup building on it should expect the question to come up in due diligence or an acquisition.
• Using it through Mistral's API is a commercial service, not a weights licence, and is governed by the API terms instead.

Devstral Small 2 and the Vibe CLI are both Apache 2.0, with no revenue clause. Open weights vs open source explains why "open" in a model announcement needs reading carefully, and this is a clear example.

The honest benchmark picture

Mistral's published figures for Devstral 2:

• SWE-bench Verified — 72.2%
• SWE-bench Multilingual — 61.3%
• Terminal-Bench 2 — 32.6%

Devstral Small 2 is listed at 68.0% on SWE-bench Verified, which is a narrow gap for a model about a fifth the size.

Mistral also ran a human evaluation through the Cline agent with an independent annotation provider. Against DeepSeek V3.2, Devstral 2 won 42.8% of comparisons and lost 28.6%. Against Claude Sonnet 4.5, Mistral says plainly that Sonnet was still clearly preferred. That candour is useful: it places Devstral 2 as a strong open option of its generation, not a match for the closed frontier at the time. Mistral's separate claim of being up to seven times more cost-efficient than Sonnet comes without a published method, so treat it as a hypothesis to test.

The Terminal-Bench figure deserves a second look. It is low next to the SWE-bench result, which suggests the model is stronger at repository patching than at long terminal-driven sessions. Agentic coding models compared covers why those two benchmarks often disagree.

Devstral 2 or Devstral Small 2

A short decision rule:

• Pick Small 2 if your company is above the revenue threshold, if you want to self-host on a single GPU, or if you need image input for screenshots and mock-ups. Apache 2.0 removes the legal question entirely.
• Pick Devstral 2 if you are under the threshold or using the API, and your own tests show the extra four points of SWE-bench translating into fewer failed runs on your codebase.

Mistral recommends a temperature of 0.2 for Devstral 2. Run both sizes on the same twenty tasks from your history, at that setting, through the same agent. Compare cost per completed task rather than pass rate alone; at a quarter of the output price, Small 2 can lose a few tasks and still come out ahead.