[Leo-rq-yu]

EmpirioLabs AI serves chat, Responses, Messages, the Gemini API, images, embeddings, speech and decisions as their specs have them, with no vendor for it: each format is configured by hand, its images answer as jobs, and its speech at a URL. Thus, add an empiriolabs vendor that lists each model by the endpoint that serves it, with its limits, features and price, and lays a dialect over two formats: its images are asked to wait and answer bytes, and its speech is called at /audio/speech:stream, whose events are read whole as the audio. An image of no output_format is now of the type its bytes say, as its answers name none.

Its transcriptions and videos are left out: a transcription is always a job, which no openai-transcriptions client could wait on, and a video is only at a signed URL off its API, which a client of OpenAI Videos could never reach. Its reranks, OCR, search, 3D and music have no kind.

Checked live through the gateway: /v1/models (133 models of 5 kinds); Chat Completions, Responses whole and streamed, Messages and generateContent (mistral-small-3-1); embeddings (text-embedding-v4); decisions (aplomb-1); images whole, streamed and edited (flux-2-klein-4b); speech as audio and as events, and a speed refused (qwen-audio-3-0-tts).

Refs: #21

[Adam-Dalloul]

Thanks for adding EmpirioLabs AI, Ruiqing! I checked it against our API: the base URL, EMPIRIOLABS_API_KEY and the model list (read live from /v1/models, so new models appear without a release) are all correct. One small ask: could the user-facing name read "EmpirioLabs AI" in the vendors table, README and docs? Happy to help with the live check.

[Leo-rq-yu]

Thanks for checking it against your API, Adam! The vendors table, README and docs now read "EmpirioLabs AI". We'd gladly take you up on the live check: a run of the chat, Responses, Messages, embeddings and video formats against your API would let us tick that box.

[Adam-Dalloul]

Thanks, Ruiqing. I ran the live check against our API at 8820fbc, built from the branch's Dockerfile with only EMPIRIOLABS_API_KEY set. Chat Completions, Responses, Messages and generateContent (each whole and streamed), embeddings, decisions, images, speech and transcriptions all answered 200, and /v1/models listed 172 of our models across all seven kinds. Models used: qwen3-8-flash, text-embedding-v4, aplomb-1, flux-2-klein-4b (one 256x256 image), and qwen-audio-3-0-tts for a short clip that whisper-large-v3-turbo transcribed word for word. A hunyuan-video-1-5 job with seconds: 1 and size: 640x360 reached us as 480p, 16:9, 1 s and polled to completed.

Four things to fix:

1. The finished video can't be fetched: /v1/videos/{id}/content answers 400. Our video URLs are signed and download without the API key, so registering this provider with LocateContent and fetching the job's URL with no key would serve it.
2. The aspect ratio goes out in lowest terms, so 854x480 becomes 427:240, and models that check the ratio fail the job. Snapping to the nearest standard ratio (16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, 21:9) would fix it.
3. A failed job's error.code is the HTTP status as a number, and error_of in videos.rs and failed in transcriptions.rs read it only as a string, so clients see failed.
4. Speech refuses any speed, but our stream takes it for most speech models (qwen-audio-3-0-tts streamed fine at 1.2) and answers 400 itself where a model can't, so that check can go.

Small one: in the README's Chat row, "EmpirioLabs AI" landed after the sentence, so it reads "whichever of the four formats it speaks EmpirioLabs AI". Happy to run it again after an update.