Recuperar un modelo
Recupera una instancia de modelo, proporcionando información básica sobre el modelo, como el propietario y las capacidades soportadas.
Autorizaciones
Gateway auth: send Authorization: Bearer <mka1-api-key>. For multi-user server-side integrations, you can also send X-On-Behalf-Of: <external-user-id>.
Encabezados
Optional external end-user identifier forwarded by the API gateway.
Parámetros de ruta
Respuesta
Está bien
-9007199254740991 <= x <= 9007199254740991llm, embeddings, images, speech_to_text, text_to_speech Modalities this model accepts as input, mirroring the registry's capabilities.modalities.input (e.g. ["text", "image"]). Clients use it to decide whether an attachment — an image, audio clip, or video — can be sent to this model at all.
Modalities this model can produce, mirroring the registry's capabilities.modalities.output. Unconstrained strings rather than a closed enum so a new registry modality reaches clients without a breaking schema change.
Context window in tokens, mirroring the registry's contextWindow. For an LLM it is the input plus output tokens one request can span: the Responses API trims the oldest conversation items to fit it, reserving nothing for the output, unless the request sets truncation to disabled, which rejects an oversized request instead; a separate request-size limit in bytes applies the same way. Chat completions forwards requests unchanged. For an embeddings model /embeddings derives its per-input character cap from it, which GET /embeddings/models reports as max_input_length; a gateway-wide per-input ceiling also applies. On other model types the gateway does not read it, so a recorded value is informational only and not enforced. Null when the registry does not record one.
-9007199254740991 <= x <= 9007199254740991Largest input image declared for this model, mirroring the registry's capabilities.maxImageResolution. A recorded bound does not mean the model accepts images; input_modalities decides that. Only the Responses API reads it, scaling a larger image down to fit before forwarding it and keeping its aspect ratio; no other endpoint, chat completions included, resizes for it. When a minimum is also set it is applied last, so an image with an extreme aspect ratio can still end up beyond this bound. Null when the registry does not record a limit.
Smallest input image declared for this model, mirroring the registry's capabilities.minImageResolution. A recorded bound does not mean the model accepts images; input_modalities decides that. Only the Responses API reads it, scaling a smaller image up to meet it before forwarding it and keeping its aspect ratio; no other endpoint, chat completions included, resizes for it. Null when the registry does not record a limit.
Output vector sizes declared for this model, mirroring the registry's embeddingLimits.dimensions, which is normally recorded only on embeddings models. A model with a single fixed size reports it as min, max and default alike, so when min equals max the size is fixed and a request should omit dimensions. Vector stores reject a size outside the range. /embeddings does not check it: it drops a requested dimensions for a fixed-size model and otherwise forwards it to the upstream. Null when the registry does not record dimensions; a vector store using such a model must then be created with an explicit size (embedding_dimensions on POST /vector_stores). Distinct from the vector store's own embedding_dimensions, which is a single size.
Total parameter count of the weights behind this model, mirroring the registry's modelCard.parameterCount. Null when the registry does not record one (closed-weight API models, or entries nobody has filled in yet).
-9007199254740991 <= x <= 9007199254740991Parameters active per token for mixture-of-experts models, mirroring the registry's modelCard.activeParameterCount. Null for dense models and whenever unrecorded.
-9007199254740991 <= x <= 9007199254740991Public release date of the model as an ISO calendar date (YYYY-MM-DD), mirroring the registry's modelCard.releaseDate. Null when unrecorded.
Curated public generation capabilities for this model.