Devstral Small 2 is the December 2025 small Devstral, currently served free on Mistral's labs endpoint with a 256K window. It is listed at $0 / $0 per Mtok, which makes Devstral Small 2 a zero-cost path into a real coding-agent model rather than a demo. Inside Atlas, it is the model to reach for when you want to run the full agent loop without a bill, either through the labs endpoint or locally at roughly 14GB quantized. The catch is that the free labs endpoint carries no availability guarantee.
Is Devstral Small 2 really free to use with Atlas?
Yes. Devstral Small 2 is listed at $0 / $0 per Mtok on Mistral's labs endpoint, making it a zero-cost path into a real coding-agent model. Set `MISTRAL_API_KEY` and select labs-devstral-small-2512 from the Atlas `/models` dialog to drive a full agent session at no cost.
Free models usually mean a crippled context window or a chat-only interface. Devstral Small 2 is neither: the December 2025 small Devstral carries a 256K window and is genuinely open weights, so the $0 / $0 listing is the same model you would otherwise host yourself. Inside Atlas that means the whole loop is available at zero cost, including hybrid semantic and keyword retrieval fused by reciprocal rank fusion, tool calls through Model Context Protocol servers, and unified diffs surfaced for approval before writing.
How well does Devstral Small 2 perform in an agent harness?
Devstral Small 2 scores near the top of the local-model pack on SWE-Bench Verified for its size, which is why it holds up inside agent harnesses instead of just autocompleting. That benchmark measures real repository tasks, which is the same shape of work Atlas asks a model to do.
The distinction between a model that autocompletes and a model that operates is the whole question for a coding agent. Devstral Small 2 lands on the operating side, and SWE-Bench Verified is the evidence: it scores near the top of the local-model pack for its size. Inside Atlas, that translates into a model that can be trusted with the loop, since Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and it computes a unified diff for every file edit and surfaces it for approval before writing.
What hardware does Devstral Small 2 need to run locally?
Devstral Small 2 is 24B parameters at roughly 14GB quantized, so it fits a 16GB GPU or an M-series Mac with headroom. Run `ollama pull devstral` to pull the 24B agent model, then register an ollama provider block in atlas.json to point Atlas at it.
The 14GB footprint is what makes Devstral Small 2 practical on hardware people actually own. Add the ollama provider block to atlas.json with baseURL `http://localhost:11434/v1` and `"npm": "@ai-sdk/openai-compatible"`, then run `atlas models ollama` to confirm Atlas sees it before selecting it in `/models`. Because Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, a local Devstral Small 2 gives you an Atlas loop where neither the index nor the generation leaves your machine.
Can you use the free Devstral Small 2 labs endpoint in production?
No. The free labs endpoint serving Devstral Small 2 carries no availability guarantee, so treat it as an evaluation channel, not production. Use labs-devstral-small-2512 to decide whether the model is worth your hardware, then move to a local `ollama pull devstral` deployment for anything you depend on.
Free with no guarantee is a fair trade for evaluation and a bad one for a workflow you rely on daily. The clean sequence for Devstral Small 2 in Atlas is: set `MISTRAL_API_KEY`, select labs-devstral-small-2512, run real tasks against your repository, and judge the diffs. If Devstral Small 2 holds up, pull the 24B weights locally at roughly 14GB and stop depending on the labs endpoint entirely. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keeping both configured costs nothing.
When should you pick a different model instead of Devstral Small 2?
Pick a different model than Devstral Small 2 on unfamiliar, poorly documented code, where it is meaningfully weaker than Devstral 2 or any frontier model. Devstral Small 2 is a 24B model with a 256K window, and 24B parameters have limits that a $0 / $0 price cannot argue away.
Devstral Small 2 is strong for its size and free, and both of those framings matter. On code it has good signal about, it operates well inside the Atlas agent loop. On unfamiliar, poorly documented code, where the model has to infer intent from thin evidence, it falls behind Devstral 2 and every frontier tier. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, which limits the damage of a wrong edit, but the better move is to escalate the session rather than repair it.
Setup
- 01Hosted: set `MISTRAL_API_KEY` and select labs-devstral-small-2512 from the `/models` dialog.
- 02Local: run `ollama pull devstral`, which pulls the 24B agent model at roughly 14GB.
- 03Add the ollama provider block to atlas.json with baseURL `http://localhost:11434/v1` and `"npm": "@ai-sdk/openai-compatible"`.
- 04Run `atlas models ollama` to confirm Atlas sees it, then pick it in `/models`.
- 05Treat the free labs endpoint as an evaluation channel, since it carries no availability guarantee.
Frequently asked questions
- Is there a free model I can use with a terminal coding agent?
- Devstral Small 2 is listed at $0 / $0 per Mtok on Mistral's labs endpoint. Set `MISTRAL_API_KEY` in Atlas and select labs-devstral-small-2512 from `/models`.
- How do I run Devstral Small 2 locally?
- Run `ollama pull devstral` to pull the 24B agent model at roughly 14GB, add an ollama provider block to atlas.json with baseURL `http://localhost:11434/v1`, then run `atlas models ollama` to confirm Atlas sees it.
- How much VRAM does Devstral Small 2 need?
- Devstral Small 2 is 24B parameters at roughly 14GB quantized, so it fits a 16GB GPU or an M-series Mac with headroom.
- What is the Devstral Small 2 context window?
- Devstral Small 2 has a 256K window, and it is the December 2025 small Devstral from Mistral.
- Is Devstral Small 2 good enough for real coding tasks?
- Devstral Small 2 scores near the top of the local-model pack on SWE-Bench Verified for its size, which is why it holds up inside agent harnesses instead of just autocompleting.
- Can I rely on the free Mistral labs endpoint?
- No. The free labs endpoint carries no availability guarantee, so treat Devstral Small 2 there as an evaluation channel and move to a local `ollama pull devstral` deployment for anything you depend on.
- What is the difference between Devstral Small 2 and Devstral 2?
- Devstral Small 2 is the 24B open-weights model on a 256K window at $0 / $0 on Mistral's labs endpoint. It is meaningfully weaker than Devstral 2 on unfamiliar, poorly documented code.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Onboard to an Unfamiliar Codebase with Atlas in 2026
How to onboard to an unfamiliar codebase with Atlas in 2026: use codebase_search, glob, read, lsp, task, and todowrite to build a mental model fast.
Atlas for Deno: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Deno. Tighten --allow-all down to real permission scopes, add JSR deps, and run deno test and deno check in 2026.
Atlas for Quarkus in 2026
Atlas is a terminal-native AI coding agent for Quarkus in 2026. It reads CDI beans and JAX-RS resources, then runs ./mvnw test behind a permission prompt.
Atlas for Scala in 2026
Atlas is a terminal-native AI coding agent for Scala in 2026. Run it in a project with a build.sbt, let it read your traits and implicits, and approve every diff.
Atlas vs Fine.dev: Terminal AI Coding Agents in 2026
Compare Atlas and Fine.dev for terminal AI coding in 2026. Atlas offers a free core TUI with permission-gated tools, while Fine.dev provides asynchronous cloud agents and prebuilt libraries.
Atlas vs GitHub Copilot CLI: A Developer's Guide for 2026
Comparing Atlas and GitHub Copilot CLI for developers in 2026. Explore pricing, privacy, agentic workflows, and extensibility to choose the best terminal AI coding agent.
Atlas for Next.js in 2026
Adopt Atlas, the terminal-native AI coding agent, for Next.js development in 2026. Enhance productivity across App Router, server components, and API routes with secure, reviewable AI assistance.
Atlas vs Base44: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent, with Base44, the Wix-owned no-code app builder, for developers in 2026. Evaluate features, pricing, and workflow.