Concepts

Models

Tiers

Every model is in one of three tiers. The router chooses a tier first, then a model within it.

Tier Typical use
small Classification, extraction, rewriting, short answers
medium Multi-step work, moderate code, analysis
large Long reasoning, large code changes, research-grade work

Variants

Each model family is served as a pair:

Variant Id Behaviour
standard utmost/llama-4-scout The upstream checkpoint, with its own refusals. Any tier can call it.
abliterated utmost/llama-4-scout-abl The same weights with the refusal direction removed. Needs a Verified organization.

An abliterated model's base_model field names its standard sibling. In the launch catalog every pair shares the same context window, output limit and capabilities, but check each model's fields rather than assuming it.

Abliteration removes trained refusals. It does not remove the hard limits, which the gateway enforces on every request.

Eval cards

Every model has an eval card. A model is not served, listed or routable until its card is published. That holds for our models and for bring-your-own models.

A card shows:

  • Refusal rate: refused prompts out of the total, on a named prompt set, and the languages tested.
  • Capability: per benchmark, the base model's score, this checkpoint's score, the difference, and the number of items.
  • Over-refusal and quality notes: written observations, for example where the standard model refuses benign requests.
  • Method: for abliterated checkpoints, the tool, technique, layers and strength used.
  • Hashes: of the base weights and the served checkpoint, so you can confirm what you are calling.
  • License: of the served checkpoint.

GET /v1/models links each card in eval_card_url. When eval_card_sample is true, the card holds sample numbers rather than measured results. Treat those as placeholders.

Pinning

Pass a model id to pin the request to that model:

{ "model": "utmost/deepseek-v4.1-flash", "messages": [{ "role": "user", "content": "..." }] }

A pinned request:

  • runs on that model only. If every deployment of it is unavailable, it fails with model_unavailable (503). It is never sent to a different model.
  • gets no refusal retry.
  • fails with context_length_exceeded (400) when the estimated input is larger than the model's context window.

Use an auto alias when you want the router to choose, fall back and retry.

Limits applied per model

  • max_tokens and max_completion_tokens above the model's max_output_tokens are lowered to it without an error.
  • reasoning_effort is dropped for models without reasoning support. See Reasoning.

Model list

Model Tier Variant Context Input /1M Cached /1M Output /1M

Prices are USD per 1M tokens. Batch requests cost half. See Billing.

GET /v1/models

No key is needed. With a key, the list also includes your organization's private models, and available and rate_limit reflect your tier.

curl https://api.utmost.site/v1/models -H "Authorization: Bearer $UTMOST_API_KEY"

The list starts with the three router aliases:

{ "id": "auto", "object": "model", "created": 0, "owned_by": "utmost", "router": true,
  "description": "Routes each request to the cheapest model likely to succeed." }

Then one object per model:

{
  "id": "utmost/llama-4-scout-abl",
  "object": "model",
  "created": 1795066216,
  "owned_by": "utmost",
  "display_name": "Llama 4 Scout · abliterated",
  "family": "Llama",
  "tier": "medium",
  "variant": "abliterated",
  "base_model": "utmost/llama-4-scout",
  "context_window": 131072,
  "max_output_tokens": 16384,
  "pricing": { "currency": "USD", "unit": "1M tokens", "input": 0.115, "cached_input": 0.0115, "output": 0.46, "batch_discount": 0.5 },
  "rate_limit": { "requests_per_minute": 600, "note": "Per key; a key's own limit overrides this." },
  "capabilities": { "tools": true, "reasoning": false, "json_mode": true, "streaming": true },
  "license": "Llama 4 Community License",
  "eval_card_url": "https://utmost.site/models/utmost/llama-4-scout-abl",
  "eval_card_sample": false,
  "available": true
}
Field Meaning
owned_by utmost, or you for your private models
tier, variant, base_model See above
context_window, max_output_tokens In tokens
pricing USD per 1M tokens; batch_discount is the multiplier for batch requests
rate_limit.requests_per_minute Your tier's default per key. Without a key, the Developer default.
capabilities tools, reasoning, json_mode, streaming
available Whether your organization can call it. Without a key, true only for standard models.

GET /v1/models/{id} returns one object. Model ids contain a slash; put it in the path as is:

curl https://api.utmost.site/v1/models/utmost/llama-4-scout-abl

An unknown id returns model_not_found (404).