Devstral 2: Mistral's Open Agentic Coding Model

Devstral hardware requirements: what PC or workstation runs it?

Devstral Small 2 needs a 24 GB graphics card to run at Q4_K_M or better; a 16 GB card can hold the 3-bit files; 8 GB and 12 GB cards cannot hold a usable quantisation entirely in VRAM. Devstral 2 (123B) needs roughly 48 GB for the smallest sensible file and 96 GB for Q4_K_M. Mistral's own guidance is that the 24B is “light enough to run on a single RTX 4090 or a Mac with 32GB RAM” and that the 123B “requires a minimum of 4 H100-class GPUs” for deployment.

How the “memory needed” column is worked out: GGUF file size, plus the key-value cache for a 16K-token context, plus 1 GB for compute buffers. The cache is arithmetic from each model's params.json: layers × KV heads × head dimension × 2 (keys and values) × 2 bytes. For Small 2 that is 40 × 8 × 128 × 2 × 2 = 163,840 bytes per token, or 2.7 GB at 16K. For the 123B it is 88 layers, so 360,448 bytes per token, or 5.9 GB at 16K. File sizes are decimal gigabytes from the Small 2 GGUF repository and the 123B GGUF repository.

Devstral Small 2 (24B)

• 8 GB card: no. The smallest published file (IQ2_XXS) is 6.55 GB, and with an 8K cache and buffer that is 8.9 GB. Ollama will split the model between GPU and system RAM instead, which runs but is slower.
• 12 GB card: only the 3-bit IQ3_XXS file at about 8K context, or a larger file split with system RAM.
• 16 GB card: Q3_K_M at 16K context. Q4_K_M (18.0 GB needed) does not fit.
• 24 GB card: Q4_K_M up to about 32K context, Q5_K_M at 16K.
• The full 256K window costs 42.9 GB of cache on its own at f16, so on a single desktop card you set a much shorter context than the model supports.

Devstral 2 (123B)

The 123B is a dense model, so unlike a Mixture-of-Experts model every weight is used for every token and there is no cheap way to park part of it in system RAM. If your build is one 24 GB card, compare the other models that fit it in best local models for a 24 GB GPU; how the quantisation names trade size for quality is covered in quantization explained.

Release date and current status

There was no Devstral release in June 2026. This page was first published that month, which is why some searches pair the two; the model itself dates from December 2025.

Mistral's model documentation for Devstral 2 and Devstral Small 2 now reads “This model is deprecated. Use Mistral Medium 3.5 for new integrations.” That is about Mistral's hosted API. For running locally the picture is different: Medium 3.5 is a 128B dense model and Mistral Small 4 has 119B total parameters, so neither is a single-card download, while Devstral Small 2 at 24B is. The site's Mistral Medium 3.5 page covers the successor.

Specs at a glance

Parameter totals are the safetensors counts reported by the Hugging Face API. Layer counts are from each repository's params.json. Tool and image support are from the model cards and match the capability badges on Ollama.

Devstral 2 benchmarks: what Mistral reports

These figures are Mistral's, copied from the table on both model cards. This site has not reproduced them. SWE-bench Verified measures whether a model, driven by an agent scaffold, can resolve real GitHub issues.

The announcement also describes a human evaluation run by an independent annotation provider with tasks scaffolded through Cline: Devstral 2 had a 42.8% win rate against a 28.6% loss rate versus DeepSeek V3.2, and Mistral states that Claude Sonnet 4.5 “remains significantly preferred”. Mistral later reported 77.6% on SWE-Bench Verified for Medium 3.5, the model that replaced Devstral 2 in its own tools.

Sources: Devstral 2 model card, Devstral 2 announcement, Medium 3.5 announcement. Scores from these benchmarks depend on the agent scaffold, so numbers from different publishers are not directly comparable.

Licensing: the two models differ

Read the licence file itself before you ship anything built on the 123B. This summary is not legal advice.

Ollama tags and install

The plain devstral name on Ollama is not Devstral 2. It resolves to 24b-small-2505-q4_K_M, the first-generation model from May 2025. Devstral 2 lives under two separate library names.

# Devstral Small 2, q4_K_M, 15 GB download
ollama pull devstral-small-2:24b
ollama run devstral-small-2:24b

Ollama publishes only three quantisations per model. For the files in between, such as Q5_K_M or the 3-bit files for a 16 GB card, use the GGUF repositories linked in the hardware section with llama.cpp. The model card notes that community GGUF files need a recent llama.cpp build and links the specific change.

Both Devstral 2 models carry Ollama's tools badge, so they work with agents that depend on tool calls. The Cline + Ollama setup guide covers wiring a local model into an editor agent, and Ollama tool calling lists which models support it. Tag list sources: devstral-small-2, devstral-2, devstral.

Mistral Vibe CLI

Mistral shipped Vibe CLI with Devstral 2. The announcement describes it as an open-source command-line coding assistant, released under Apache 2.0, that works in a terminal or inside an editor through the Agent Communication Protocol, and it is available as a Zed extension. The model card gives the install command as uv tool install mistral-vibe. The source is at github.com/mistralai/mistral-vibe.

On first run Vibe asks for a Mistral API key, and Mistral has since made Medium 3.5 its default model. To use it with a local Devstral you have to point it at your own endpoint; check the repository for the current configuration options.

Which one to pick

This section is opinion, based on the sizes and licences above.

Frequently asked questions

Sources

• Mistral AI: Introducing Devstral 2 and Mistral Vibe CLI (December 9, 2025)
• Devstral Small 2 model card and Devstral 2 model card
• Devstral 2 modified MIT licence
• Ollama devstral-small-2 tags, devstral-2 tags
• Mistral model documentation: Devstral 2 (deprecation notice)