ModelService
List models
List the models available through Inworld Inference, with their capabilities, pricing, and specifications.
/llm/v1alpha/modelsBrowse inworld.ai/models without an API key. To get model metadata through this endpoint, use Authorization: Basic $INWORLD_API_KEY. Requests without credentials return 401.
Use this endpoint to list models. The OpenAI-style GET https://api.inworld.ai/v1/models (client.models.list() in the OpenAI SDK) is not available.
The optional provider and model query parameters filter by exact match, for example ?provider=openai. For an Inworld-hosted model, combine model with provider=inworld.
By default, the list contains models you can call with chat completions. Add ?output_modalities=image for image models, ?output_modalities=video for video models, ?output_modalities=decisions for decision models or ?output_modalities=extractions for extraction models.
Each model's id is the value to send as model in a request, such as inworld/models/<name> for a model Inworld hosts; send it unchanged.
Each model's endpoints lists the request paths that serve it. To list only the models an API serves, send that path, such as ?endpoints=/v1/responses. Image models served by the Images API list /v1/images/generations; to list only those, send ?output_modalities=image&endpoints=/v1/images/generations.
A video model's entry carries spec.video, the values POST /v1/videos accepts for it (lengths, resolutions, aspect ratios, frames and reference limits), and pricing.video, its per-second prices. See Model compatibility.
Return only models from this provider. Must match the model's provider value exactly (for example, openai).
Return only the model with this exact model value (for example, gpt-5.4). Combine with provider to select a single provider's listing. For an Inworld-hosted model, set provider=inworld.
Return only models whose spec.outputModalities includes one of these values: text, image, video, decisions or extractions. Separate several values with commas, or pass all to list every model. Without it, the response lists the models you can call with chat completions; video models, decision models, extraction models and models served by the Images API are listed only when you ask for them (for example, ?output_modalities=image).
Return only models that every listed request path serves, spelled as in a model's endpoints, such as /v1/images/generations. Separate several paths with commas. It applies together with output_modalities: to list Images API models, send ?output_modalities=image&endpoints=/v1/images/generations. A path that takes no model returns 400.
The list of available models.
Show child attributes
The model's name at its provider. Models hosted by Inworld are listed as models/<name>. To call the model, send id as model in a request rather than this field.
The organization that created the model. May differ from the provider (e.g., a Meta model hosted on Groq).
Pricing information for the model (in USD). Text models carry per-token prices; video models carry video, and Images API models carry image, instead. Fields without a price are omitted.
Show child attributes
A video model's prices, in USD. Set only for video models.
Show child attributes
Price per second of generated video, one entry per resolution the model offers.
Price per input image (a frame or reference image). Omitted when the model doesn't charge for input images.
Technical specifications and capabilities of the model.
Show child attributes
The input modalities supported by the model. file means PDF documents. For decision models, image, file, audio and video are the input types the Decisions API accepts in state.
outputModalitiesenum<string>[]
The output modalities supported by the model. Decision models report decisions.
The maximum number of tokens the model can process as input (context window size).
The generation parameters supported by this model. Common values include: max_tokens, temperature, top_p, stop, seed, response_format, structured_outputs, tools, tool_choice, reasoning, include_reasoning, frequency_penalty, presence_penalty.
High-level capability flags for the model. Only capabilities that are true are included in the response.
Show child attributes
true confirms the model supports native web search (the provider's own search grounding, web_search_options or web_search with engine: native). false doesn't rule it out: some models report false and still search. Router-run web search works on any model with tool calling.
Whether the model supports structured output with a response schema (JSON mode).
Reasoning support in detail. Present when the model supports reasoning.
responseFormatCapabilityobject
Structured output support in detail. Present when the model supports JSON output.
Show child attributes
Whether the catalog lists JSON mode (response_format: {"type": "json_object"}) for the model. It isn't a guarantee: on some models that report true, the format is dropped and the reply is ordinary text, so parse the reply before you rely on it.
For a video model: the values POST /v1/videos accepts for it. The API checks each request against these limits. Set only for video models.
Show child attributes
Every clip length the model makes (seconds), in whole seconds, ascending. A range lists each second of it.
The output resolutions the model makes (resolution), lowest first.
Show child attributes
The longest prompt the model takes, in characters (Unicode code points). Omitted when the model sets no maximum.
The shortest prompt the model takes, in characters (Unicode code points). Omitted when the model sets no minimum.
supportedFrameImagesenum<string>[]
The frame_images frame types the model takes. Empty or omitted when it takes no frames. Frames can't be combined with references.
The reference media (input_references) a request may carry. Omitted when the model takes no references.
Show child attributes
One entry per reference type the model takes. A type that isn't listed is rejected.
Show child attributes
The reference type, as input_references[].type names it.
Available options:image_urlvideo_urlaudio_url
The shortest a single reference of this type may be, in seconds. Omitted for images, or when the model sets no minimum.
The longest a single reference of this type may be, in seconds. Omitted for images, or when the model sets no maximum.
How many references a request may carry in total. Omitted when only the per-type limits apply.
Types at least one reference must be of. For example, ["image_url", "video_url"] means audio clips can't be sent on their own. Empty or omitted when the model requires none.
How the model uses an audio_url reference. speech: the model voices the clip itself where the prompt quotes the words the clip says, so a person in the video can say your recording; a prompt that quotes other words gets those words in the clip's voice. voice: the model takes only the clip's voice and generates the speech itself, saying the words the prompt quotes in that voice; it never plays the clip, whatever the clip says. Omitted when the model takes no audio references or names neither use. Exact timing is not guaranteed in either case.
Available options:speechvoice
Indicates whether this model is currently supported and recommended for use. Omitted when false; such models may be experimental, deprecated, or not yet generally available.
Recent reliability and latency statistics for the model. Omitted when no data is available.
Show child attributes
Average request latency in milliseconds. Returned as a string (64-bit integer JSON encoding).
The request paths that serve this model, such as /v1/chat/completions, /v1/messages, /v1/responses, /v1/videos, /v1/images/generations or /v1/images/edits. A path not listed rejects the model. Use it to choose the API to call.