Ministral 3B is the cheapest model Mistral sells: $0.04 / 1M input tokens and $0.04 / 1M output tokens, with a 128,000 token window. It exists for the boring high-volume calls, which in Atlas means session titles, commit messages, and summarization. Ministral 3B is far too small to drive an agentic coding loop, so it belongs in the small_model slot only, with a real coding model left in the primary slot.
What is Ministral 3B for in an Atlas session?
Ministral 3B exists for the boring high-volume calls, which in Atlas means session titles, commit messages, and summarization. At $0.04 / 1M input tokens and $0.04 / 1M output tokens it is the cheapest model Mistral sells, so those calls stop showing up on the invoice at all.
Every Atlas session produces a stream of calls nobody thinks about. Naming the session. Summarizing a turn that ran long. Drafting a commit message after a change lands. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so commit message generation alone is a recurring cost. Ministral 3B absorbs that traffic. It is a 3B-class edge model, and it will not impress anyone on a hard task, but a session title does not need to be impressive. It needs to be instant and nearly free, and Ministral 3B is both.
How much cheaper is Ministral 3B than Mistral Small 3.2?
Ministral 3B is roughly 12x cheaper on input than Mistral Small 3.2. At $0.04 / Mtok symmetric, Ministral 3B charges the same for output as for input, so the cheap slot in Atlas costs about as much as rounding error on a normal working day.
Symmetric $0.04 / Mtok pricing is what makes Ministral 3B different from the rest of the lineup, where output typically costs three or four times input. For Atlas's small_model slot that symmetry matters, because summarization is an output-heavy job: you feed it a long turn and get back a paragraph, over and over. On a model with a 4x output multiplier, that pattern is where the money goes. On Ministral 3B, it does not. Compare with Mistral Small 3.2 at $0.10 / Mtok input before deciding, because the cheaper model is only a win if it does the job.
Why can Ministral 3B not drive an Atlas coding loop?
Ministral 3B is far too small to drive an agentic coding loop: it will lose the thread on multi-file edits. Despite a 128,000 token window, a 3B-class model does not hold a plan across a long tool sequence, so Atlas will produce confident and wrong diffs.
Context window size and reasoning capacity are different things, and Ministral 3B is the clearest example of the gap. It can read 128,000 tokens and 128,000 tokens of output are available to it, but the model behind that window is a 3B edge model. Multi-file work needs the agent to remember what it changed two tool calls ago and why. Ministral 3B does not. Tool-call reliability is also weak compared to Devstral, which is the other reason to keep it out of the primary model slot. Leave `"model"` on a real coding model such as mistral/devstral-medium-2507.
How does Ministral 3B keep latency low in the Atlas TUI?
Ministral 3B is designed for on-device and edge inference, so latency in Atlas's small_model slot is very low. Because titles and summaries fire constantly in the TUI, first-token speed on a $0.04 / Mtok model is felt more than any capability difference would be.
Atlas is a terminal-native TUI rendered with SolidJS through the OpenTUI renderer, and the small_model slot is on the interactive path. A slow title generation is a visible stall. Ministral 3B's edge-oriented design means it returns quickly, which keeps the interface feeling responsive even when the primary model is chewing through a large refactor. That responsiveness, not the price, is often what convinces people to keep Ministral 3B in the cheap slot after they have tried a larger model there.
When should you upgrade from Ministral 3B to a bigger small_model?
Upgrade from Ministral 3B when Atlas's cheap slot has to call tools. Ministral 3B has weak tool-call reliability compared to Devstral, so once subagents start doing real work rather than writing titles, a 3B edge model becomes the bottleneck.
The upgrade signal is concrete: subagents misfiring on tool schemas. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and if those subagents run on `"small_model"`, the cheap model is suddenly in the tool loop. Ministral 3B is not built for that. At that point, either move subagent work back to the primary model or step the cheap slot up. If you never see tool-call failures, the $0.04 / 1M input tokens price is doing exactly what it should and there is no reason to pay more.
Setup
- 01Export MISTRAL_API_KEY before launching Atlas.
- 02Run `atlas models mistral` and look for ministral-3b-latest.
- 03In atlas.json, set only the cheap slot: `"small_model": "mistral/ministral-3b-latest"`.
- 04Leave `"model"` on a real coding model such as mistral/devstral-medium-2507 and let Atlas route the cheap traffic to Ministral.
- 05Watch for subagent tool-call failures: they are the signal that the cheap slot needs a larger model.
Frequently asked questions
- what is the cheapest mistral model
- Ministral 3B is the cheapest model Mistral sells, at $0.04 / 1M input tokens and $0.04 / 1M output tokens, with a 128,000 token context window. It is built for high-volume calls like titles and summaries.
- how do I set ministral 3b as small_model in atlas
- Export MISTRAL_API_KEY, run `atlas models mistral` and look for ministral-3b-latest, then in atlas.json set only the cheap slot: `"small_model": "mistral/ministral-3b-latest"`, leaving `"model"` on a real coding model.
- can ministral 3b write code in a coding agent
- Not as the primary model. Ministral 3B is far too small to drive an agentic coding loop and will lose the thread on multi-file edits. Leave `"model"` on a coding model such as mistral/devstral-medium-2507.
- what is the context window of ministral 3b
- Ministral 3B has a 128,000 token context window and a 128,000 token output budget, which is generous for a 3B-class edge model. Window size does not compensate for its limited reasoning capacity on multi-file work.
- ministral 3b vs mistral small 3.2 for the cheap slot
- Ministral 3B at $0.04 / Mtok is roughly 12x cheaper on input than Mistral Small 3.2. Small 3.2 is the safer choice when the cheap slot has to call tools, since Ministral 3B has weak tool-call reliability.
- why is my agent generating bad session titles
- If Atlas's small_model is Ministral 3B, it is a 3B-class edge model chosen for cost and latency, not depth. That is the right trade for titles and commit messages, but move anything requiring judgement to the primary model slot.
- is ministral 3b good for on device inference
- Yes. Ministral 3B is designed for on-device and edge inference, which is why latency in Atlas's small_model slot is very low. It returns quickly on the interactive path in the TUI.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Python in 2026
Atlas is a terminal-native AI coding agent for Python in 2026. Run it in a repo with a pyproject.toml or requirements.txt and review every diff before it lands.
Run Atlas Headless in CI with Atlas (2026 Workflow)
How to run Atlas headless in CI in 2026: atlas run sends one prompt and exits when the session goes idle, with --format json, --command, and --continue for pipeline steps.
Atlas for SolidJS: A Terminal-Native AI Coding Agent Built on Solid, for Solid, in 2026
Atlas is a terminal-native AI coding agent for SolidJS in 2026. It reads your createSignal and createMemo graph, fixes destructured props that break reactivity, and runs vitest.
Atlas for React Native: Terminal-Native AI Coding Across the Native Boundary in 2026
Atlas is a terminal-native AI coding agent for React Native in 2026. Work across the New Architecture, native modules, and platform-specific files with diff-first review.
Atlas for SQL in 2026
Atlas is a terminal-native AI coding agent for SQL in 2026. Run it in a repo with your migrations or .sql files, optimize a query, and review the diff before applying.
Research a Third-Party API Before Integrating It with Atlas in 2026
How to research a third-party API with Atlas in 2026: websearch finds the current docs, webfetch pulls the page as markdown or text, and grep checks repo conventions.
Atlas for dbt: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for dbt. Read the ref() DAG, convert a table model to incremental, run dbt build against dev, and add tests in 2026.
Atlas vs Ellipsis: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Ellipsis, a cloud platform with usage-based pricing and live session tracing for 2026.