Llama 3.3 8B Instruct (Meta Llama API) is the small member of the 3.3 family on Meta's hosted API, and it belongs in Atlas's small_model slot. Llama 3.3 8B Instruct keeps the full 128,000 token window of its 70B sibling, so Atlas's cheap slot can still read a big file without truncation. Pricing is not publicly listed, the models.dev entry for Meta's Llama API records $0 / $0. The 4,096 token output ceiling rules it out as a primary code-writing model.
What is the small_model slot in Atlas and why use Llama 3.3 8B?
Atlas's small_model slot handles the cheap work: titles, summaries, and context reading that does not need a frontier model. Llama 3.3 8B Instruct (Meta Llama API) is a good fit because it keeps the full 128,000 token window of its 70B sibling, so cheap does not mean truncated.
Most small models pay for their size with a small window, which defeats the purpose: the cheap slot ends up unable to read the file it was asked to summarize. Llama 3.3 8B Instruct (Meta Llama API) does not make that trade. Set `"small_model": "llama/llama-3.3-8b-instruct"` in atlas.json and keep a stronger model in `"model"`. Atlas lets you switch the active model and provider on the fly with favorites and recents, and the small_model slot is separate from that switching, so it keeps serving background work no matter what the primary model is.
Why is Llama 3.3 8B Instruct not a primary coding model?
Llama 3.3 8B Instruct (Meta Llama API) has a 4,096 token output ceiling, which rules it out as a primary code-writing model in Atlas. An 8B dense model will also drop instructions on long agentic tool chains, so keep Llama 3.3 8B Instruct in the small_model slot.
Two failure modes stack. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and diffs are output tokens, so a 4,096 token cap truncates anything beyond a modest edit. Separately, Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and both agents chain tool calls. An 8B dense model loses the thread on long chains. Neither problem is dangerous, because every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, but both make the model unproductive as the driver.
How do you configure Llama 3.3 8B Instruct in atlas.json?
Export LLAMA_API_KEY, confirm llama-3.3-8b-instruct with `atlas models llama`, then in atlas.json set `"small_model": "llama/llama-3.3-8b-instruct"` and keep a stronger model in `"model"`. Llama 3.3 8B Instruct is served over the OpenAI-compatible endpoint at api.llama.com, which Atlas wires with @ai-sdk/openai-compatible.
The configuration for Llama 3.3 8B Instruct (Meta Llama API) is two lines because Meta's endpoint speaks the OpenAI-compatible protocol Atlas already understands. No bespoke adapter, no custom provider code, just a key and two model ids in the same provider block. Verify with `atlas models llama` before you rely on it, since a small_model that fails to resolve produces confusing symptoms: session titles stop generating while the primary model keeps working. Resolution failures are quiet in the small slot, which is exactly why you check it explicitly.
How much does Llama 3.3 8B Instruct cost on the Meta Llama API?
Pricing for Llama 3.3 8B Instruct (Meta Llama API) is not publicly listed. The models.dev entry for Meta's Llama API records $0 / $0, which means an unpublished rate rather than a guarantee, so budget for a small_model whose cost you cannot yet quantify.
The cost uncertainty is more tolerable in the small_model slot than in the primary slot, which is another argument for putting Llama 3.3 8B Instruct (Meta Llama API) there. Background work in Atlas, session titles and summaries, is a small fraction of total tokens even at a 128,000 token window, so an unpublished rate on the cheap slot is a bounded risk. If the rate turns out badly, the escape is quick: the identical open weights run on a laptop, so only the provider id in atlas.json changes.
Can you run the same 8B weights offline as a fallback?
Yes. Llama 3.3 8B Instruct is small enough that the identical open weights run on a laptop, giving you an offline fallback with the same prompts. If you later need offline parity, pull the equivalent open weights into Ollama and only the provider id in atlas.json changes.
Prompt parity is the underrated part. Because Llama 3.3 8B Instruct (Meta Llama API) and the local build are the same weights, the summaries and titles Atlas generates do not shift in style or quality when you go offline. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so a local 8B in the small_model slot plus a local embedder means none of the background work touches a network either. The 4,096 token output ceiling is an endpoint property of Meta's API, not of the weights.
When should you pick a different model than Llama 3.3 8B Instruct?
Pick a different model for anything in Atlas's primary `"model"` slot. Llama 3.3 8B Instruct (Meta Llama API) has a 4,096 token output ceiling and 8B dense weights that drop instructions on long agentic tool chains, so a build agent driven by it will stall on real refactors.
The natural pairing is Llama 3.3 8B Instruct (Meta Llama API) in small_model and Llama 3.3 70B Instruct in the primary slot, since both live behind the same LLAMA_API_KEY and the same OpenAI-compatible endpoint at api.llama.com. That gives Atlas one provider block, two model ids, and a clean split between cheap context reading and real code writing. If you need a larger output budget than 4,096 tokens for either slot, the Meta Llama API is the wrong endpoint and a different provider is the answer.
Setup
- 01Export LLAMA_API_KEY in your shell so Atlas can authenticate against Meta's Llama API.
- 02Confirm llama-3.3-8b-instruct with `atlas models llama`.
- 03In atlas.json set `"small_model": "llama/llama-3.3-8b-instruct"` and keep a stronger model in `"model"`.
- 04Leave the provider on @ai-sdk/openai-compatible, since Llama 3.3 8B Instruct is served over the OpenAI-compatible endpoint at api.llama.com.
- 05If you later need offline parity, pull the equivalent open weights into Ollama and only the provider id in atlas.json changes.
Frequently asked questions
- what is atlas small_model and which model should i put there
- Atlas's small_model slot handles cheap background work like titles and summaries. Llama 3.3 8B Instruct (Meta Llama API) fits because it keeps the full 128,000 token window of its 70B sibling, so the cheap slot can still read a big file.
- what is the context window of llama 3.3 8b instruct
- Llama 3.3 8B Instruct (Meta Llama API) has a 128,000 token context window, the same as Llama 3.3 70B. Output is capped at 4,096 tokens on the same endpoint.
- can llama 3.3 8b write code in atlas
- Not as the primary model. The 4,096 token output ceiling rules it out for code writing, and 8B dense will drop instructions on long agentic tool chains. Keep it in the small_model slot.
- how do i set small_model in atlas.json
- Export LLAMA_API_KEY, confirm llama-3.3-8b-instruct with `atlas models llama`, then set `"small_model": "llama/llama-3.3-8b-instruct"` in atlas.json while keeping a stronger model in `"model"`.
- how much does llama 3.3 8b cost on meta's api
- Pricing is not publicly listed. The models.dev entry for Meta's Llama API records $0 / $0, so treat the rate as unpublished rather than free when budgeting an Atlas workload.
- can i run llama 3.3 8b offline with atlas
- Yes. The identical open weights run on a laptop, so pull them into Ollama and only the provider id in atlas.json changes. Prompts and behavior stay the same.
- llama 3.3 8b vs llama 3.3 70b in atlas
- Both sit behind the same LLAMA_API_KEY and the same OpenAI-compatible endpoint at api.llama.com. Put Llama 3.3 8B Instruct in small_model for cheap context reading and Llama 3.3 70B Instruct in the primary model slot for code writing.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
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 Factory AI: Terminal AI Coding Agents in 2026
Compare Atlas and Factory AI, two terminal AI coding agents for developers in 2026. Explore features, pricing, privacy, and workflow to choose the best fit.
Atlas vs Zed: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Zed, the GPU-accelerated collaborative editor, for developers in 2026. Evaluate their AI autonomy, privacy, and workflow.
Atlas for Elixir in 2026
Adopt Atlas, the terminal-native AI coding agent, for Elixir development in 2026. Enhance productivity with deep code understanding, safety features, and direct integration into mix projects and OTP applications.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.
Debug a Single Failing Test with Atlas in 2026
How to debug one failing test with Atlas in 2026: run it in isolation with bash, walk the call graph with the lsp tool, and fix the code, not the assertion.
Atlas vs Kilo Code: Terminal AI Coding Agents in 2026
Atlas and Kilo Code in 2026: Compare terminal-native TUI vs VS Code/JetBrains agents. Evaluate pricing, code safety, deployment, and model routing for AI coding.
Atlas vs OpenHands: Terminal AI Coding Agents in 2026
Atlas and OpenHands comparison for 2026. Explore terminal-native TUI, self-hosting, code indexing, change review, and pricing models for AI coding agents.