Models

GET /v1/models is the source of truth — call it before you hardcode an id anywhere, because the table below is a snapshot of that same call and this page can go stale between deploys the way any documentation can.

GET/v1/models

Response

{
  "object": "list",
  "data": [
    {"id": "larsa-auto", "object": "model", "owned_by": "openai"},
    {"id": "larsa-general", "object": "model", "owned_by": "openai"},
    {"id": "larsa-general-fast", "object": "model", "owned_by": "openai"},
    {"id": "larsa-law-ir", "object": "model", "owned_by": "openai"},
    {"id": "larsa-law-es", "object": "model", "owned_by": "openai"},
    {"id": "larsa-stt", "object": "model", "owned_by": "openai"},
    {"id": "larsa-tts", "object": "model", "owned_by": "openai"}
  ]
}

owned_by: "openai" is a compatibility artifact of the OpenAI wire format, not a claim about who trained the weights — every id above runs on this platform's own hardware. See Introduction.

Renamed from lardad-*

Every id above used to carry the lardad-* prefix. The old names still resolve — lardad-auto, lardad-general, lardad-general-fast, lardad-law-ir, lardad-law-es, lardad-stt and lardad-tts all call the matching larsa-* model above — but treat them as deprecated aliases, not as ids to write new code against; they may be removed in a later release.

Reference

Model	For	Context	Modalities	Tools
`larsa-auto`	One id for everything — classifies your message from its text and content and dispatches to the right specialist below. See Chat.	131,072	text, image → text	Yes
`larsa-general`	The default chat and vision model — llama.cpp serving Qwen3.6-35B-A3B. Reads images and documents. See Chat and Vision.	131,072	text, image → text	Yes
`larsa-general-fast`	The same weights (NVFP4 quantisation) served by vLLM instead of llama.cpp — an alternative engine kept for comparison, not a smaller model.	131,072	text, image → text	Yes
`larsa-law-ir`	Iranian law, retrieval-grounded over 117,499 provisions before the model answers.	131,072	text → text	No
`larsa-law-es`	Spanish and EU law (BOE consolidated), retrieval-grounded over 835,951 provisions.	131,072	text → text	No
`larsa-stt`	Speech to text — Whisper large-v3, silence-based chunking for long audio. See Speech to text and Realtime transcription.	—	audio → text	—
`larsa-tts`	Text to speech — Aava for Persian, Chatterbox for 23 other languages, with a round-trip check against `larsa-stt` available on request. See Text to speech.	—	text → audio	—

Two ceilings that apply to all of them

Context, 131,072 tokens. Both chat backends were started with the same window, and the retrieval models generate on the same weights after their search step — so it is one number across the board, not a per-model detail to look up.
Output, 8,192 tokens per request, platform-wide. A generation with no ceiling can run to the context limit and hold a GPU slot for everyone else — max_tokens may ask for less than 8,192, never more. Requesting more is clamped, not refused.

The public price list at /pricing also carries a handful of rows — gpt-4o, unknown, the wildcard * — used internally for cost estimation. They are not callable model ids; GET /v1/models is the list that matters.

Reference#

Two ceilings that apply to all of them#

Reference

Two ceilings that apply to all of them