Llama 3.3 70B Instruct (Meta Llama API) is Meta's own hosted endpoint for Llama 3.3 70B, reached through the OpenAI-compatible base at api.llama.com. Llama 3.3 70B was the model that matched Llama 3.1 405B on most instruction benchmarks at a fraction of the serving cost, and on Meta's API it carries a 128,000 token context window. Pricing is not publicly listed, the models.dev entry for Meta's Llama API records $0 / $0. Max output is 4,096 tokens, which shapes how you use it in Atlas.
What is Llama 3.3 70B Instruct (Meta Llama API) best at inside Atlas?
Llama 3.3 70B Instruct (Meta Llama API) is best at reading and reasoning over a large codebase in Atlas. The 70B dense weights deliver near-405B instruction quality, which is why Llama 3.3 70B displaced Llama 3.1 405B for most deployments, and the endpoint carries a 128,000 token window.
Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, not blind line windows. A 128,000 token window on Meta's own API gives that retrieval somewhere to land. Where Llama 3.3 70B Instruct (Meta Llama API) is strongest is the reading half of an Atlas session: understanding a service, tracing a bug across files, drafting a plan. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and the plan agent is where the 70B model earns its place.
How does the 4,096 token output ceiling change Atlas workflows?
Max output on the Llama 3.3 70B Instruct (Meta Llama API) endpoint is 4,096 tokens, so Atlas must split any sizeable refactor across turns. A whole-module rewrite will not fit in one response, and the correct pattern is Atlas's plan agent followed by targeted per-file edits.
The 4,096 token cap is the defining constraint of Llama 3.3 70B Instruct (Meta Llama API) in an agent. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a diff is output tokens. Ask for a sweeping rewrite and the model runs out of room mid-diff. Ask for one file at a time and the same model performs well. Structure the session accordingly: plan first, then a sequence of small, reviewable edits. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, which makes the many-small-edits pattern cheap to iterate on.
How much does Llama 3.3 70B Instruct cost on Meta's API?
Pricing for Llama 3.3 70B Instruct (Meta Llama API) is not publicly listed. The models.dev entry for Meta's Llama API records $0 / $0, which reflects an unpublished rate rather than a guarantee of free inference, so treat cost as an open question when planning an Atlas budget.
The absence of a published rate matters when you are choosing a model to drive Atlas all day. Llama 3.3 70B Instruct (Meta Llama API) is a credible primary model on capability, and the cost side is simply not something this page can quantify. The mitigating fact is that the same open weights are available to self-host, so if Meta's pricing turns out not to fit your budget, the escape hatch is running the weights yourself rather than rewriting your workflow. Atlas lets you switch the active model and provider on the fly with favorites and recents, which keeps that option one keystroke away.
How does Atlas connect to api.llama.com?
Llama 3.3 70B Instruct (Meta Llama API) is reached through an OpenAI-compatible base at https://api.llama.com/compat/v1/, so Atlas loads it with @ai-sdk/openai-compatible and no bespoke adapter. Export LLAMA_API_KEY, then run `atlas models llama` to confirm llama-3.3-70b-instruct resolves from the registry.
The OpenAI-compatible surface is the practical reason Llama 3.3 70B Instruct (Meta Llama API) is easy to adopt. Atlas does not need provider-specific code for Meta, so wiring the model is a key plus a model id: pin `"model": "llama/llama-3.3-70b-instruct"` in atlas.json and start a session. Every Atlas tool call is still permission-gated against allow, ask, and deny rules before it runs, and every capability Atlas has, Model Context Protocol servers, plugins, subagents, git staging, works identically behind Meta's endpoint as behind any other provider.
When should you pick a different model than Llama 3.3 70B Instruct?
Pick a different model when long-context retrieval is the job. Llama 3.3 70B predates Meta's mixture-of-experts generation and is behind Llama 4 Scout on long-context retrieval, so filling the 128,000 token window with retrieved code is where Llama 3.3 70B Instruct (Meta Llama API) shows its age.
Two constraints push you off Llama 3.3 70B Instruct (Meta Llama API). The 4,096 token output ceiling rules out large single-turn rewrites, and the pre-MoE architecture costs you recall when Atlas packs a very long prompt. Neither is disqualifying for a plan-then-edit workflow, and both matter if you want one model that does everything. For teams that need reproducible open weights and near-405B instruction quality without running 405B, Llama 3.3 70B remains the pragmatic choice: it matched Llama 3.1 405B on most instruction benchmarks at a fraction of the serving cost.
Can you self-host the same Llama 3.3 70B weights?
Yes. Llama 3.3 70B Instruct (Meta Llama API) runs on Meta's endpoint, and the same open weights are available to self-host, so the identical 70B model can be moved onto hardware you control without changing the Atlas prompts or the agent behavior.
Self-hosting Llama 3.3 70B changes one line in atlas.json, the provider id, because Atlas reaches both Meta's API and a local Ollama server through OpenAI-compatible providers. What changes is the operational picture, not the workflow. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so pairing self-hosted weights with a local embedder gives a fully private setup. Keep in mind that the 4,096 token output ceiling is an endpoint property of the Meta Llama API, not a property of the weights themselves.
Setup
- 01Get a key from Meta's Llama API and export LLAMA_API_KEY.
- 02Run `atlas models llama` to confirm llama-3.3-70b-instruct resolves from the registry.
- 03Pin `"model": "llama/llama-3.3-70b-instruct"` in atlas.json.
- 04Because output is capped at 4,096 tokens, prefer Atlas's plan agent then targeted per-file edits over whole-module rewrites.
- 05Keep the OpenAI-compatible base https://api.llama.com/compat/v1/ in the provider block so Atlas loads it with @ai-sdk/openai-compatible.
Frequently asked questions
- what is the context window of llama 3.3 70b on meta's api
- Llama 3.3 70B Instruct (Meta Llama API) has a 128,000 token context window. The output ceiling on the same endpoint is 4,096 tokens per turn.
- how much does the meta llama api cost
- Pricing is not publicly listed. The models.dev entry for Meta's Llama API records $0 / $0, which reflects an unpublished rate rather than a promise of free inference.
- how do i connect atlas to api.llama.com
- Export LLAMA_API_KEY, run `atlas models llama` to confirm llama-3.3-70b-instruct resolves, then pin `"model": "llama/llama-3.3-70b-instruct"` in atlas.json. The OpenAI-compatible base is https://api.llama.com/compat/v1/.
- why does llama 3.3 70b truncate my refactor in atlas
- Max output on the Meta Llama API endpoint is 4,096 tokens. Use Atlas's plan agent, then make targeted per-file edits, rather than asking for a whole-module rewrite in one turn.
- llama 3.3 70b vs llama 3.1 405b for coding
- Llama 3.3 70B was the model that matched Llama 3.1 405B on most instruction benchmarks at a fraction of the serving cost, which is why it displaced 405B for most deployments.
- is llama 3.3 70b good at long context retrieval
- Llama 3.3 predates Meta's MoE generation and is behind Llama 4 Scout on long-context retrieval, so avoid packing the full 128,000 token window when recall accuracy matters.
- can i self-host llama 3.3 70b instead of using meta's api
- Yes. The same open weights are available to self-host, and because Atlas reaches both through OpenAI-compatible providers, only the provider id in atlas.json changes. The 4,096 token output cap is an endpoint property, not a weight property.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for OCaml: A Terminal-Native AI Coding Agent for dune and opam Projects in 2026
Atlas is a terminal-native AI coding agent for OCaml in 2026. It reads dune stanzas and .mli signatures, runs dune runtest behind a prompt, and finishes with ocamlformat.
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.
Atlas vs Sourcery: Terminal AI Coding Agents in 2026
Compare Atlas and Sourcery for 2026. Atlas offers terminal-native AI coding with permission-gated tools and local embeddings. Sourcery provides Python-first LLM review and rules-based refactoring, starting at
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 Poolside: Terminal AI Coding Agents in 2026
Compare Atlas and Poolside, two terminal AI coding agents in 2026. Atlas offers a free core and robust safety features, while Poolside provides open-weight Laguna models and on-prem deployment options.
Atlas for C in 2026
Atlas is a terminal-native AI coding agent for C in 2026. Run it in a project with a Makefile, have it find memory leaks or add Unity tests, and review the diff.
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. Goose: Choosing Your AI Coding Agent in 2026
Compare Atlas and Goose for 2026. Atlas offers terminal-native TUI and code-specialized features. Goose provides shareable Recipes and 70+ MCP extensions for general agentic workflows.