News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgfnDevstral 2

Devstral 2

Open-weight coding model family for software agents

Mistral AI released Devstral 2 and Devstral Small 2 in 2026 as open-source models for code automation. The larger 123B variant targets enterprise workflows via API, while the 24B version runs on local hardware.

Share on WhatsApp

  • Scores 72.2% on SWE-bench Verified
  • Supports 256K token context window
  • Runs locally on consumer hardware (Small 2)
  • Released under permissive open licences
  • Integrates with Mistral Vibe CLI

What it is

Devstral 2 is a family of coding-focused language models created by Mistral AI. It includes a 123-billion parameter dense transformer and a smaller 24-billion parameter variant. Both models support a context window of 256,000 tokens and are designed to handle complex software engineering tasks. The larger model aims for high performance in automated coding agents, while the smaller one enables local execution on consumer devices. They are released under permissive open-source licences to encourage widespread adoption and customization.

How it came about

Mistral AI introduced these models as the next generation of their coding tools. The release coincided with the launch of Mistral Vibe CLI, a terminal-based assistant powered by Devstral. The company positioned the launch as a stealth success, noting significant early usage metrics. This iteration builds on previous efforts to provide efficient, open-weight alternatives to proprietary coding assistants, focusing on practical deployment for developers and businesses.

Inside the model

The architecture relies on dense transformer layers rather than mixture of experts. Devstral 2 uses 123 billion parameters, whereas Devstral Small 2 uses 24 billion. Both accept text and image inputs, allowing for multimodal interactions. The models support tool calling, enabling them to execute commands and modify files. Training focused on optimizing for codebase exploration, dependency tracking, and error correction. The smaller model is optimized for single-GPU or CPU-only environments, while the larger one requires multiple high-end GPUs.

What it can do

On the SWE-bench Verified benchmark, Devstral 2 achieved a score of 72.2 percent. Devstral Small 2 scored 68.0 percent, performing comparably to models much larger than itself. In human evaluations using the Cline framework, Devstral 2 won 42.8 percent of comparisons against DeepSeek V3.2 but lost more often to Claude Sonnet 4.5. The models excel at multi-file orchestration, legacy system modernization, and bug fixing. They maintain architectural context across large projects.

How to use it

Weights are available under modified MIT and Apache 2.0 licences. The models can be accessed via the Mistral AI API, where Devstral 2 is currently free before transitioning to paid pricing. Devstral Small 2 is deployable locally on consumer hardware. Mistral Vibe CLI provides an open-source interface for terminal-based coding assistance. Integration is available for IDEs like Zed through the Agent Communication Protocol.

Reception and debate

Partners such as Kilo Code and Cline praised the model for its tool-calling reliability and smooth integration. Early adoption was strong, with over 17 billion tokens processed in the first day. Critics note that while it outperforms some open competitors, it still trails top-tier closed-source models in user preference. The compact size of the Small variant was highlighted as a benefit for privacy-focused and resource-constrained users.

Why it matters

The release reinforces the trend toward efficient, open-weight coding models. By offering a locally runnable option, it lowers barriers for individual developers and small teams. The integration with existing agent frameworks demonstrates a shift towards modular AI tools that fit into established developer workflows. It challenges the necessity of massive parameter counts for high-level coding tasks.

Questions people ask about Devstral 2

What is the difference between Devstral 2 and Devstral Small 2?

Devstral 2 has 123 billion parameters and targets data center deployment, while Devstral Small 2 has 24 billion parameters and runs on consumer hardware. Both share the same context window and multimodal capabilities.

Can I run Devstral 2 on my personal computer?

Devstral 2 requires multiple high-end GPUs. However, Devstral Small 2 is designed for single-GPU or even CPU-only setups, making it suitable for local deployment on standard consumer devices.

What licence applies to these models?

Devstral 2 uses a modified MIT licence, and Devstral Small 2 uses Apache 2.0. Both are considered permissive open-source licences that allow for commercial use and modification.

How does it perform compared to other coding models?

It scores 72.2% on SWE-bench Verified. In head-to-head tests, it beats DeepSeek V3.2 but lags behind Claude Sonnet 4.5. It offers competitive performance with significantly fewer parameters.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Devstral 2 yet, so the facts on this page come from the sources listed above: mistral.ai/news/devstral-2-vibe-cli. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.