Skip to main content
POST
Requires version 2.0.1 or later of the Archetype platform.

Overview

This endpoint takes an OpenAI Chat Completions request and sends it to a model that the Archetype Platform hosts. model names one of the available models that serves Chat Completions: newton-c-2.6 or newton-fusion-1.0. The platform forwards the body to a running copy of that model and returns its chat.completion object unchanged. You can call it with an OpenAI SDK by setting the base URL to https://platform.u1.archetypeai.app/api/inference and the API key to your Archetype API key.

Request

string
required
Name of the model to query: newton-c-2.6 or newton-fusion-1.0. A model that doesn’t serve Chat Completions, such as omega-embedding, returns 400. An unknown name returns 404.
object[]
required
The conversation, as system, developer, user, and assistant messages. A message’s content is a string or a list of parts: text, image_url, video_url, and input_audio. Tool messages and tool calls aren’t supported. Send images and video inline as data: URIs, in a request body under 2 MiB; every image and video frame is resized to 224 × 224 pixels. See Current Limitations.
boolean
default:"false"
Must be false or omitted. Streaming isn’t supported yet, and true returns 400.
integer
Maximum number of tokens to generate. max_tokens is accepted as the older name; when you set both, max_completion_tokens wins.
number
Sampling temperature, from 0 to 2.
number
Nucleus sampling cutoff, from 0 to 1.
integer
default:"1"
Number of completions. Only 1 is supported.
Options that would change the answer in a way the model can’t honor return 400: a non-zero frequency_penalty or presence_penalty, logprobs, top_logprobs, logit_bias, stop, non-empty tools or functions, a tool_choice other than none or auto, a response_format other than text, modalities other than ["text"], audio, web_search_options, and reasoning_effort. Their OpenAI defaults are accepted, so an SDK that sends every field still works.

Notes for OpenAI SDK Users

  • Newton Fusion 1.0’s reasoning arrives in the answer. The text starts with a <think>...</think> block, empty when the model answered directly; remove it before showing the answer. reasoning_effort is refused. To answer without reasoning, end the user text with /no_think: it works when the request includes an image or video, and on a text-only request the model reasons anyway. See Reasoning.
  • repetition_penalty is ignored here. It isn’t an OpenAI field; sent through the SDK (for example in extra_body), it’s accepted and has no effect. If the model repeats itself in long lists, send the request to /inference/query, which applies it.
  • Media go inline. Send images and video as data: URIs; uploaded files can’t be referenced.
  • A query stops after about 60 seconds and returns 504. Keep answers short with the token limit and the prompt; see Choosing an Endpoint.

Response

The model’s chat completion object. The answer is in choices[0].message.content.

Important Notes

  • Every error is listed in Errors on the Inference API overview.
  • Errors use the Archetype error envelope, not OpenAI’s. An OpenAI SDK raises them as API errors with the HTTP status.
  • 400 with model '<model>' refused this request means the model rejected the body, for example because it sets one of the unsupported options above.