Skip to main content

Devstral Small 2 24B

Mistral's coding-agent model, multimodal, served on NVIDIA.

Overview​

  • model_name: hf-mistralai-Devstral-Small-2-24B-Instruct-2512 (HuggingFace import, currently free)
  • Source: mistralai/Devstral-Small-2-24B-Instruct-2512
  • Architecture: Mistral3ForConditionalGeneration — 24B parameters, FP8, vision-capable
  • Served on: GPU via vLLM. 24 GB of weights fit on a single H200 or Arc Pro B70; smaller cards serve it tensor-parallel.
  • Status: ready (verified end-to-end through the public inference route, 2026-08-03)

When to use​

✅ Software-engineering tasks: reading a repository, proposing a patch, explaining a stack trace. Devstral is tuned for agentic coding rather than chat.

❌ Short conversational turns — a 7B model answers faster for the same quality. For general chat use Qwen 2.5 7B; for pure code completion see Qwen3 Coder 30B.

Live specs​

curl "https://api.colabhive.com/api/builder/v1/endpoints?visibility=public&search=Devstral"

Quick start​

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/<endpoint_id>/infer" \
  -H "X-API-Key: $COLABHIVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"messages": [{"role": "user", "content": "This test fails with a KeyError on an empty dict. Walk me through how to find the cause."}], "max_tokens": 600}}'
Weights are published twice in the source repository

The repo ships the same 24 GB in both a sharded and a consolidated layout, so it measures 48 GB on disk. Only one set is loaded — the endpoint's declared VRAM reflects the 24 GB that actually reach the GPU.

Authors: José Luis Minich, Maximiliano Lucius.