Generic OpenAI-compatible

Use Generic OpenAI-compatible when your server supports the request and response formats below and has no matching dedicated backend profile.

Obtain an endpoint

Generic is a protocol baseline, not a server program to install. Use the URL, model ID, and authentication supplied by your server operator or hosted provider. If you want to run a server yourself, start with the backend launch guides and choose its qualified profile. Check that the selected model supports the operation you want to use.

Configure a connection

1. Open Connections → Add connection, choose its purpose and Generic OpenAI-compatible profile.
2. Name it and enter the API base URL, normally ending in /v1.
3. Configure authentication. For an HTTP URL, enable Allow HTTP for this connection only if you trust the network. Choose Save connection.
4. Open the workflow’s Settings cog and choose Transcription for Voice or Audio file, Cleanup for either transcription workflow, or Speech for Text to speech. Select the connection, choose Refresh models or enter the exact model ID, then save. Model discovery reads metadata only.
5. Explicitly try one operation with the model you chose and review the result.

See Connect a speech server for HTTP permission, separate servers, and credential configuration.

Implemented capabilities

Operation	Request and response
Microphone and audio-file transcription	Multipart file, model, response_format=json, and optional language, prompt, and temperature; completed JSON with string text.
Optional file streaming	stream=true; typed transcript delta/done events or legacy per-segment text events. A server may return completed JSON instead.
Transcript cleanup	Non-streaming text chat completions with system/user string messages, temperature zero, and optional max_tokens.
Speech playback	model, input, string voice, speed, and response_format=wav; fully buffered PCM16 WAV audio.

Enter the API base, not an operation URL

Freehand appends audio/transcriptions, chat/completions, or audio/speech. A full operation URL would append the path again. Confirm your server accepts the fields and formats above; an OpenAI-compatible label alone is not enough.

Available options

Area	Limit or failure behavior
Request fields	No model-name inference or automatic provider-only parameters
File streaming	Progressive output for an uploaded file; does not enable live microphone or Realtime API sessions
Stream completion	Typed streams need a final transcript; legacy segment streams finish at EOF; incompatible streams are not resubmitted
Upload size	Server limits may be lower than Freehand’s 2 GiB ceiling; oversized stored files are not split automatically
Speech audio	Mono/stereo PCM16 WAV at 8–192 kHz, buffered up to 32 MiB; a WAV MIME type does not establish sample encoding
Cleanup failure or length limit	Falls back to raw text under normal delivery and cancellation rules

The dedicated OpenAI hosted profile is planned. Hosted and self-hosted endpoints can use Generic when the selected model accepts this exact contract. See the full protocol reference for bounds and failure handling.

Optional transcription controls

Workflow	Where to find the controls
Voice	Voice transcription → Transcription → Request settings
Audio file	Audio file → Transcription → Transcription controls

Context hint is sent as prompt. Temperature is sent only when its override is enabled. Leave either unset for server defaults; each model may interpret these fields differently. Generic appends enabled shared vocabulary to prompt; it does not send hotwords.

See Transcription controls for limits, persistence, and the difference between recognition hints and cleanup.

Cleanup generation controls

Control	Generic behavior
Output-token limit	Optional max_tokens, from 1 to 65,536; the model may impose a smaller limit
Other token-limit fields	Does not substitute max_completion_tokens, send both fields, or retry a rejection
Reasoning	No reasoning fields; turn off a custom reasoning override before switching from llama.cpp or vLLM

S1-mini needs server-side reasoning disabled

Configure that on the server, or use qualified llama.cpp or vLLM profiles for automatic S1-mini request enforcement. Leave the output limit off if your endpoint/model does not accept max_tokens.

See generation controls.

Language selection

Use the language guide to choose Server default, automatic detection, a named code, or a custom value. The selected model determines which languages work. If you enable S1-mini cleanup, review its English-only language behavior.