Baseten: Mammoth 1 API Reference

About

Baseten Mammoth 1 is a 27B model served by Baseten for typed decisions over text, structured data, and images. It supports Noul yes/no probabilities, Choice classification, and Score rubrics, returning structured answers without free-form chat generation. Each question has a 16,384-token context limit including its shared state and formatting. Multiple questions reuse the shared input prefix, and image input is supported through user messages. The platform exposes the standard Baseten route through its Decisions API.

1. Calling the API

Setup your API Key

Create an Infron account and API key.

export INFRON_KEY="YOUR_API_KEY"

Submit a request

Send the embedded red-square example to the platform endpoint:

curl --request POST "https://llm.onerouter.pro/v1/decisions" \
  --header "Authorization: Bearer $INFRON_KEY" \
  --header "Content-Type: application/json" \
  --data '{
  "model": "baseten/mammoth-1",
  "state": {
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Inspect the supplied image."
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "data:image/png;base64,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"
            }
          }
        ]
      }
    ]
  },
  "questions": {
    "square_color": {
      "type": "choice",
      "instructions": "What is the color of the square in the image?",
      "criteria": {
        "red": "Red",
        "blue": "Blue",
        "green": "Green"
      }
    }
  },
  "provider": {
    "only": [
      "baseten"
    ],
    "allow_fallbacks": false
  },
  "usage": {
    "include": true
  }
}'

Response

This is an illustrative response, with placeholder request ID, probabilities, and usage:

{
  "id": "request_id_example",
  "request_id": "request_id_example",
  "model": "baseten/mammoth-1",
  "answers": {
    "square_color": {
      "type": "choice",
      "choice": "red",
      "confidence": 0.97,
      "probabilities": {
        "red": 0.98,
        "blue": 0.01,
        "green": 0.01
      }
    }
  },
  "usage": {
    "input_tokens": 100,
    "output_tokens": 1
  }
}

Routing and usage

provider.only=["baseten"] selects Baseten; provider.service_tier="standard" selects the standard route. Platform routing options and usage.include are not sent upstream. The upstream tier is its default tier. Actual upstream input and output token counts are preserved. Shared-prefix input is counted once, followed by each question's additional input; image tokens are included. The upstream reports one output token per question, even though output is free.

Errors

Configured type and per-question option limits are checked before forwarding. The upstream enforces prompt length, the combined 512-option budget, image constraints, and conversation rules; invalid requests return an error rather than truncating the input. Authentication and capacity errors follow the existing platform error handling. This route is synchronous and does not support streaming.

2. Schema

Input

Question types

Choice accepts 1–255 named options. Score accepts 1–10 ordered levels, numbered from zero. Their descriptions may be strings, objects, arrays, or null. Noul may omit criteria, use null or an empty object, or specify exactly true and false; the upstream validates this contract. Missing, null, or empty instructions use the upstream's question defaults. Each Noul question consumes two of the combined 512 options. Each question's complete prompt, including images and formatting, must fit 16,384 tokens or a smaller max_len.

Media inputs

Use image parts in state.messages[].content for user messages, as shown above. PNG, JPEG, and WebP data URLs are accepted. HTTPS image URLs must be hosted on Amazon S3, Google Cloud Storage, or Azure Blob Storage. A request accepts at most 16 images. Image detail must be auto or omitted. Conversation roles are user and assistant; assistant images, audio, and video are unsupported. Generic JSON state is passed through without serializing it into a string.

Output

3. Detailed Pricing

Standard-tier official pricing applies to actual input usage, including image tokens. No promotional or commercial discount is configured. Output token counts remain unchanged. Official pricing; model pricing; API contract.