Skip to main content
Open WebUI is a self-hosted chat interface for local and hosted models. Its image, voice and document features can call any OpenAI-compatible API, so you can point them at deAPI and keep your chat model where it is: Ollama, OpenAI or any other connection.
deAPI does not serve /v1/chat/completions. Keep your existing chat connection and use deAPI for images, speech, transcription and embeddings.
Every step below was checked by sending deAPI the exact requests that Open WebUI v0.13 sends.

Prerequisites

Every setting below lives under Admin Panel → Settings and uses the same two values:

Image generation

  1. Open Admin Panel → Settings → Images.
  2. Turn on Image Generation.
  3. Set Image Generation Engine to Default (Open AI).
  4. Enter the API Base URL and API Key. Leave API Version empty.
  5. Type a model name into Model. The dropdown only lists OpenAI’s own models, so type the name instead.
  6. Set Image Size to a size the model supports, then click Save.
Open WebUI’s default size, 512x512, works on all four. It asks for b64_json, and deAPI returns a PNG. In testing, a 512x512 image from ZImageTurbo_INT8 came back in about 14 seconds.

Image editing

  1. On the same Images page, turn on Image Edit.
  2. Set Image Edit Engine to Default (Open AI).
  3. Enter the same API Base URL and API Key.
  4. Set Model to Flux_2_Klein_4B_BF16 and click Save.
Flux_2_Klein_4B_BF16 accepts the multipart image[] field that Open WebUI uses, combines up to 3 input images, and returned a 512x512 edit in about 8 seconds in testing. QwenImageEdit_Plus_NF4 (one input image) works too, but is much slower: minutes rather than seconds in testing. See Image editing for limits.

Text-to-speech

  1. Open Admin Panel → Settings → Audio.
  2. Under Text-to-Speech, set Text-to-Speech Engine to OpenAI.
  3. Enter the API Base URL and API Key.
  4. Set TTS Model to Kokoro and TTS Voice to a Kokoro voice such as af_nova, then click Save.
Read Aloud under a reply now plays audio generated by deAPI. A short sentence took about 3 seconds in testing.
Kokoro voices are tied to one language, set by the voice prefix, and the model does not detect the language from the text. Pick a voice that matches the language of your replies. The OpenAI names alloy, echo, fable, onyx, nova and shimmer also work on Kokoro, and all of them map to English voices. Voice lists for Kokoro, Chatterbox and Qwen3_TTS_12Hz_1_7B_CustomVoice are in Text-to-speech.

Speech-to-text

  1. On the same Audio page, set Speech-to-Text Engine to OpenAI.
  2. Enter the API Base URL and API Key.
  3. Set STT Model to WhisperLargeV3Ct2 and click Save.
Voice input is now transcribed by deAPI. In testing, WebM (what Chrome records), M4A, MP3 and WAV files all transcribed correctly; a short clip took about 5 seconds. Audio files can be up to 20 MB.

Document embeddings

  1. Open Admin Panel → Settings → Documents.
  2. Set Embedding Model Engine to OpenAI.
  3. Enter the API Base URL and API Key.
  4. Set Embedding Model to Bge_M3_FP16 and click Save.
Bge_M3_FP16 returns 1024-dimensional vectors and handles many languages.
Vectors from different embedding models are not compatible. After switching, run Reindex Knowledge and Memory Vectors on the same page so existing documents are embedded again with Bge_M3_FP16.

Troubleshooting


Next steps

OpenAI Compatibility

Every endpoint, model and field the OpenAI-compatible API supports, and where it differs from OpenAI