Devstral

weights

open(Apache-2.0, on HF)

data

closed

code

partial(inference/serving

license

Mistral-Modified-MIT(non-OSI

Devstral 2's smaller 24B model ships under Apache-2.0, but the larger 123B model carries a non-OSI Modified MIT license; since the two are sizes of one release, the record follows the more restrictive license. Training and post-training data and the pipeline behind the checkpoints are undisclosed, which keeps the family at open weights regardless of which license the SKU uses.

• https://huggingface.co/api/models?search=Devstral&sort=createdAt&direction=-1&limit=30 recorded 2026-09-29

  Hub search for Devstral, newest first: no first-party "mistralai/Devstral" repo newer than the 2512 pair appears; newer entries are community quantizations, GGUF conversions and LoRAs such as "Devstral-Small-2-24B-Instruct-2512-GGUF".

• https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512 recorded 2026-09-29

  Model page for the 123B flagship: the tag list includes "license:other", and the page text reads "This model is licensed under a" followed by a link to the Modified MIT License file.

• https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512/raw/main/LICENSE recorded 2026-09-29

  LICENSE file, "Modified MIT License": "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million" - a revenue ceiling, so not OSI.

• https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512/raw/main/README.md recorded 2026-09-29

  Model card README: "Apache 2.0 License**: Open-source license allowing usage and modification for both commercial and non-commercial purposes." and "This model is licensed under the [Apache 2.0 License]". Searched the body for training data, dataset and corpus: none; only inference and serving code is referenced (vLLM, mistral_common, the Mistral Vibe CLI), no training code or post-training data.

Devstral is a specialist coding-agent model, adopted mainly through OpenHands and similar coding-agent tooling rather than general chat use. Adoption is measured as Hugging Face downloads summed across the two declared Devstral 2 checkpoints, most of it on the smaller 24B model rather than the 123B flagship.

• https://huggingface.co/api/models/mistralai/Devstral-2-123B-Instruct-2512 recorded 2026-08-14

  32,695 downloads in the trailing 30 days for mistralai/Devstral-2-123B-Instruct-2512; 333 likes; gated false

• https://huggingface.co/api/models/mistralai/Devstral-Small-2-24B-Instruct-2512 recorded 2026-08-14

  188,901 downloads in the trailing 30 days for mistralai/Devstral-Small-2-24B-Instruct-2512; 641 likes; gated false

• https://huggingface.co/mistralai/Devstral-Small-2507 recorded 2026-08-13

  97,969 downloads in the trailing 30 days on the superseded SKU the earlier reading came from

Devstral 2's 123B model scores 72.2% on SWE-bench Verified on its model card, ahead of Qwen3 Coder Plus, Kimi K2 Thinking and GLM 4.6, while the smaller 24B model scores 68.0%. DeepSeek v3.2 and the proprietary models on the same table score higher, and Mistral's newer Medium 3.5 model has replaced Devstral 2 as its coding model, which keeps Devstral a step below the frontier.

• https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512/raw/main/README.md recorded 2026-09-29

  Benchmark Results table in the model card: the "Devstral 2" row reads "72.2%" on SWE Bench Verified, "61.3%" on SWE Bench Multilingual and "32.6%" on Terminal Bench 2; "Devstral Small 2" reads "68.0%", "55.7%" and "22.5%"; competitor rows include "Kimi K2 Thinking", "GLM 4.6" and "DeepSeek v3.2", with "Claude Sonnet 4.5" at "77.2%"; the footnote reads "Benchmark results presented are based on publicly reported values for competitor models".

• https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/raw/main/README.md recorded 2026-09-29

  Vendor card for a newer model: "Mistral Medium 3.5 supersedes all our previous coding models, namely Devstral, across all benchmarks" and "It scores **91.4%** on τ³-Telecom and **77.6%** on SWE-Bench Verified."