Llama 3.2 3B (Ollama) is a 2.0GB model with a 128K token (131,072) context. That is the entire pitch: an edge-class model that still holds a long transcript. Inside Atlas the pricing is Free (self-hosted), and roughly 4GB to serve means it runs on a laptop with integrated graphics or no GPU at all, which makes it the realistic floor for an Atlas setup. Use it as small_model so Atlas titles, summaries, and cheap subagent calls stay local and free. A 3B will not carry a multi-file refactor, and pretending otherwise wastes your afternoon.
Can you run Atlas with no GPU at all?
Yes. Llama 3.2 3B (Ollama) is a 2.0GB download that needs roughly 4GB to serve, so it runs on a laptop with integrated graphics. Llama 3.2 3B is the realistic floor for an Atlas setup with no GPU at all, and the pricing is Free (self-hosted).
The no-GPU case is real and usually ignored. A developer on a work-issued laptop with integrated graphics, no API budget, and a policy against sending source code to a vendor has, on paper, no path to an AI coding agent. Llama 3.2 3B is that path. Run ollama pull llama3.2:3b (2.0GB, aliased to :latest), register it in the atlas.json ollama models map with limit.context 131072 and limit.output 8192, and Atlas has a working model. Llama 3.2 3B is free and offline, which makes it viable on an air-gapped machine where no API key exists at all. The capability ceiling is low, and the page is explicit about that below, but the floor is a functioning agent.
How does a 3B model hold a 128K context window?
Llama 3.2 3B (Ollama) carries a full 128K token (131,072) context in a 2.0GB download, which is unusual for an edge-class model. Meta distilled Llama 3.2 3B from the larger Llama 3.1 models, so instruction following is unusually good for 3B and the long window comes with it.
Parameter count and context length are separate axes, and Llama 3.2 3B is the clearest demonstration of that. The 128K window means Llama 3.2 3B can hold a long Atlas transcript, including retrieved code chunks and the unified diffs Atlas computes for every file edit, without dropping the early turns. What 3B parameters cannot do is reason deeply over all of it. So the useful mental model is a model with excellent memory and modest judgment. That combination is a good fit for summarization, titling, and triage work, which is precisely the small_model role, and a bad fit for authoring a nontrivial patch.
What should Llama 3.2 3B actually do inside Atlas?
Set small_model to ollama/llama3.2:3b so Atlas titles, summaries, and cheap subagent calls stay local and free. Llama 3.2 3B belongs in small_model, not in the build slot, because a 3B will not carry a multi-file refactor no matter how large its 128K window is.
Atlas generates a steady stream of low-stakes text: session titles, summaries, and the cheap calls that Atlas subagents make when Atlas fans out work into foreground or parallel background sessions. Those calls do not need a frontier model, and routing them to one is the most common source of a surprising bill. Llama 3.2 3B does that work at Free (self-hosted) pricing on 4GB of memory, alongside whatever larger model you run in the build slot. On a machine that can only hold one model, Llama 3.2 3B can be the whole setup, but scope your expectations to single-file, well-scoped edits.
Is a 3B model safe to give write access to your repository?
Keep Atlas permissions on ask when running Llama 3.2 3B (Ollama), since a 3B model should not be chaining unsupervised writes. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so the guardrail is a configuration setting, not a hope.
Model size and permission posture should move together. A strong model on a well-understood repository can reasonably run with more allow rules. Llama 3.2 3B is a 2.0GB distilled model, and its judgment is the weakest link in the loop, so the correct posture is ask on anything that writes. Atlas backs that up structurally: Atlas computes a unified diff for every file edit and surfaces it for approval before writing, Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back. With those three together, a weak model is contained rather than dangerous.
When should you pick a different model than Llama 3.2 3B?
Pick a different model than Llama 3.2 3B (Ollama) as soon as you have the memory for one. Llama 3.2 3B will not carry a multi-file refactor, and it is text only, with no vision, unlike the Llama 3.2 vision variants that are not in this tag family.
Llama 3.2 3B is a floor, not a destination. If a machine can serve 7GB, Llama 3.1 8B brings the same 128K window with considerably more capability. If it can serve 16GB, a 24B general model or a code specialist changes what Atlas can attempt in a single session. The reason to stay on Llama 3.2 3B is a hard hardware or network constraint: integrated graphics, an air-gapped machine, or a policy that forbids API keys. Atlas lets you switch the active model and provider on the fly with favorites and recents, so upgrading later is a pull and a line in atlas.json, not a migration.
Setup
- 01Pull it: ollama pull llama3.2:3b (2.0GB, aliased to :latest)
- 02Register llama3.2:3b in the atlas.json ollama models map with limit.context 131072 and limit.output 8192
- 03Set small_model to ollama/llama3.2:3b so Atlas titles, summaries, and cheap subagent calls stay local and free
- 04Keep Atlas permissions on ask, since a 3B model should not be chaining unsupervised writes
- 05Run the codebase index with the Ollama embedder so even a CPU-only machine gets full hybrid semantic and keyword search without a cloud account
Frequently asked questions
- can I run an AI coding agent without a GPU
- Yes. Llama 3.2 3B is a 2.0GB model that needs roughly 4GB to serve, so it runs on a laptop with integrated graphics. It is the realistic floor for an Atlas setup with no GPU at all.
- what is the smallest model that works with atlas
- Llama 3.2 3B at 2.0GB is a practical floor. It holds a 128K token (131,072) context and is Free (self-hosted), but it belongs in small_model rather than the build slot.
- does llama 3.2 3b support vision
- No. Llama 3.2 3B is text only, with no vision, unlike the Llama 3.2 vision variants that are not in this tag family.
- how do I set small_model in atlas.json
- Set small_model to ollama/llama3.2:3b so Atlas titles, summaries, and cheap subagent calls stay local and free, after registering the tag with limit.context 131072 and limit.output 8192.
- can llama 3.2 3b do a multi-file refactor
- No. A 3B will not carry a multi-file refactor even with a 128K window. Keep Llama 3.2 3B on titles, summaries, and cheap subagent calls, and use a larger model in the build slot.
- can atlas run on an air-gapped machine
- Yes. Llama 3.2 3B is free and offline, and running the codebase index with the Ollama embedder gives even a CPU-only machine full hybrid semantic and keyword search without a cloud account.
- what permissions should I set for a small local model
- Keep Atlas permissions on ask when running Llama 3.2 3B, since a 3B model should not be chaining unsupervised writes. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Java in 2026
Adopt Atlas, the terminal-native AI coding agent, for Java development in 2026. Enhance your workflow with intelligent code search, refactoring, and robust safety features for Maven and Gradle projects.
Atlas vs PearAI: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with PearAI, a VS Code fork. Evaluate features, pricing, and community support for developers in 2026.
Atlas for JavaScript in 2026
In 2026, Atlas empowers JavaScript developers with a terminal-native AI coding agent. It indexes code by AST, uses local embeddings, and offers permission-gated tools for safe, efficient development.
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 Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
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 vs Codebuff: Terminal AI Coding Agents in 2026
Atlas and Codebuff are terminal AI coding agents for 2026. Compare Atlas's terminal-native TUI, permission-gated tools, and diff review with Codebuff's multi-agent system and flexible pricing.
Atlas for Rust in 2026
Adopt Atlas, the terminal-native AI coding agent, for Rust development in 2026. Tackle borrow checker errors and clippy lints with Atlas's secure, approval-gated assistance.