AI services

TranscribeX uses an AI service for titles, summaries, chat, AI translation and dictation polish. You can keep everything on your Mac with the built-in Local LLM or Apple Intelligence, or connect a provider with your own API key. Manage them in Settings › AI Service.

Built-in options

Two services are always in the list and need no API key.

Local LLM

A language model that runs on your Mac, even offline. Set it up in Settings › Local LLM:

1. Under Available Models, download a model:
   • Qwen3.5 4B (Recommended): about 3.06 GB, 262k context, uses about 2 GB of memory while running.
   • Gemma3 1B: about 0.8 GB, 32k context, uses about 1 GB. Choose it if your Mac has less memory.
2. Pick it in Default Model.
3. Optionally try it in Test Model: choose the model, type a Question and click Ask.

Other settings:

• Preload Model On Launch (on by default) loads the default model when the app starts, so the first request is faster.
• Timeout Interval (5 mins by default) stops a request that produces no new output for that long. Model loading isn’t counted. Long transcripts need more time, so raise it if summaries of long meetings stop early.
• Model Storage Path sets where models are stored. Delete downloaded models before changing it.

Apple Intelligence

Uses Apple’s on-device model. It “Works only on macOS 26 with Apple Intelligence enabled in System Settings.” In per-job chat, only the first 8,000 characters of a transcript are sent to Apple Intelligence, so use another service for long recordings.

Add a provider

1. Open Settings › AI Service.
2. Choose a Provider and click Add.
3. Click the new card to expand it and fill in the fields.

Every card also has:

• Instance Name: the name shown in menus. You can add the same provider more than once, for example with different models.
• Default: makes this the default service. Only one service can be default.
• Timeout: 1 min to 30 mins (1 min by default). Raise it for slow models or long transcripts.

For Ollama and LMStudio, start the server app first, then click the refresh button next to Model Name to load your models.

Custom (OpenAI-compatible) endpoints

Custom works with any service that offers an OpenAI-compatible chat completions API. Enter the base URL (for example one ending in /v1); TranscribeX adds /chat/completions if it’s missing. The API key is sent as a bearer token. There’s no dedicated Anthropic (Claude) provider. If a provider you want to use offers an OpenAI-compatible endpoint, add it as Custom.

Which features use which service

Note: On a new install Apple Intelligence is the default service. If your Mac isn’t on macOS 26 with Apple Intelligence turned on, make Local LLM or a provider the default, or titles and summaries won’t generate.

Privacy: what is sent where

• Local LLM and Apple Intelligence run on your Mac. Nothing leaves it.
• Ollama and LMStudio send text to the server URL you set, which is usually your own Mac or local network.
• OpenAI, Google, DeepSeek, Perplexity and Custom send the transcript text (or, for global Chat, the relevant passages) and your prompt to that provider, under your API key and their terms. Audio is never sent to AI services.

Transcription itself is separate. Local transcription models never upload audio; see choosing a model.