Atlas Inference
An OpenAI-compatible API for open-weight text models. Change a base URL, a key, and a model name — nothing else.
Atlas serves a curated ladder of open-weight text models behind an OpenAI-compatible HTTP API. A client already written against an OpenAI SDK works against Atlas after three changes and no new code:
https://api.inference.runatlas.com/v1API key
An Atlas key, beginning atl_live_. Created in the console, shown once.
Model
An Atlas catalog identifier, such as atlas-mid-1.
Your first request
There is no Atlas SDK, and there is not meant to be one. The OpenAI Python and Node clients are the supported path — a client library of our own would be a second integration surface to keep in step with the first.
Start here
Quickstart
Get a key and make a request, streaming and non-streaming.
Model catalog
The four models, their context lengths, capabilities, and provisional prices.
OpenAI compatibility
Exactly which fields are honored, which are accepted and ignored, and what is not offered at all.
Error codes
The status codes and error.type values, and how to tell a rate limit from
an overload.
What the platform gives you
Streaming
Server-sent events in the OpenAI chunk format, terminated by [DONE].
Tool calling
The tools / tool_choice request shape your client already emits.
Structured output
response_format: json_schema, validated against a published subset.
Rate limits
Per-organization, per-model request and token budgets, with standard headers.
Usage and cost
Per-model, per-key, and over-time breakdowns, priced at the provisional rates.
Zero retention
Prompt and completion content is not retained past the response.
Beta terms
Atlas is in an invite-only beta. Usage is free, and the published prices are provisional — they are what your usage would cost, shown so you can model it, and they may change before billing is enabled. See Pricing.
There is no public sign-up. Organizations are created by Atlas from an invite; see Organizations and roles.