Models

A model id is always vendor/name — google/gemma-4-26b-a4b, lgai/exaone-4.0-32b. There are no variant suffixes: no :free, no :nitro, no :floor. One id means one model, and how it is served is a routing decision described in Provider routing.

Where models come from

Every model in the catalog is served by one or both of these:

When a model exists in both places, the local slot wins by default. That is the point of the service; see Local-first routing.

The models you can use today

These are public and available to any account:

google/gemma-4-26b-a4b is the default model — it is what you get if you omit model and models entirely.

Listing models

curl https://openrouter.myip.co.kr/api/v1/models

No authentication required. The response is {"data":[…]}, one object per model:

{
  "id": "lgai/exaone-4.0-32b",
  "canonical_slug": "lgai/exaone-4.0-32b",
  "name": "LG AI: EXAONE 4.0 32B",
  "created": 1786000000,
  "description": "EXAONE 4.0 32B (AWQ-INT4). MyIP local GPU slot.",
  "context_length": 32768,
  "architecture": {
    "modality": "text->text",
    "input_modalities": ["text"],
    "output_modalities": ["text"],
    "tokenizer": "Other",
    "instruct_type": null
  },
  "pricing": {
    "prompt": "0.000030000000",
    "completion": "0.000150000000",
    "request": "0",
    "image": "0",
    "web_search": "0",
    "internal_reasoning": "0",
    "input_cache_read": "0.000003000000",
    "input_cache_write": null,
    "currency": "KRW"
  },
  "top_provider": { "context_length": 32768, "max_completion_tokens": 32768, "is_moderated": false },
  "supported_parameters": ["max_tokens", "temperature", "top_p", "top_k", "stop", "seed",
    "frequency_penalty", "presence_penalty", "repetition_penalty", "logit_bias", "response_format"],
  "per_request_limits": null,
  "links": { "details": "/api/v1/models/lgai/exaone-4.0-32b/endpoints" }
}

Reading pricing

Prices are KRW per token, encoded as strings — the same encoding the OpenAI-compatible ecosystem uses, with a different unit. "0.000030000000" means ₩30 per 1M prompt tokens.

currency is a field we added, and X-MyIP-Currency: KRW is on the response. Both exist so that a client written against a USD-denominated API cannot quietly misread the number. request, image and web_search are always "0": we do not have per-request, image, or search billing.

Reading supported_parameters

This is the list of request parameters the model advertises. It is advisory — we forward every parameter we do not consume, and the upstream decides what to do with the rest. If you need a hard guarantee, use provider: { "require_parameters": true } so that candidates missing the parameters you sent are dropped from the chain.

Note what is present and absent: response_format is there on our local models (Structured outputs), and tools/tool_choice only on models that actually support them (Tool calling).

Filtering the catalog

GET /api/v1/models accepts these query parameters:

Examples:

# Models that accept images
curl "https://openrouter.myip.co.kr/api/v1/models?input_modalities=text,image"

# Models with tool calling, cheapest first
curl "https://openrouter.myip.co.kr/api/v1/models?supported_parameters=tools,tool_choice&sort=price"

# Models served from our own GPUs
curl "https://openrouter.myip.co.kr/api/v1/models?providers=local-gpu"

Related endpoints

• GET /api/v1/models/count → {"data":{"total":N}}, the number of public, enabled models.
• GET /api/v1/models/user (Bearer, inference key) → models this key can actually call. It includes models that are not in the public list but excludes models with no working provider mapping.
• GET /api/v1/models/{author}/{slug}/endpoints → every endpoint serving that model, external and local, with context, quantization, and a status field. Local slots appear here with tag set to the engine (vllm or llamacpp), and status: 0 while the slot is running.
• GET /api/v1/providers → the provider list, including the slugs you use in provider.order / only / ignore.

curl https://openrouter.myip.co.kr/api/v1/models/google/gemma-4-26b-a4b/endpoints

When a model disappears

Models can be disabled by our operators, and provider mappings can be turned off. The catalog is the truth at any moment; a request for a disabled or unknown id returns 400 model_not_found. A known id with no usable candidate left — every provider disabled, filtered out by your provider{} preferences, or in cooldown — returns 404 no_endpoints_found instead. The two are different codes because the fixes are different.