Skip to main content
Anarlog does not hide one model behind the whole app. It keeps two independent selections:
  • Transcription turns meeting audio into text.
  • Intelligence turns text into summaries, titles, and chat responses.
Open Settings → Transcription and Settings → Intelligence to see the selection active on your computer. Changing one selection does not change the other. For managed Pro (Cloud) and Auto selections, this page explains the current upstream route.

​
How a meeting moves through AI

Recording, editing, and local storage do not require an AI model. A hosted Transcription provider receives audio. A hosted Intelligence provider receives the text and other context needed for that request.

​
Anarlog Cloud

Anarlog Cloud uses managed routes. Settings displays the managed selection as Pro (Cloud). Transcription uses that Cloud route. Intelligence offers Auto, displayed as Pro (Cloud), or a curated model choice. Auto lets Anarlog choose a suitable upstream model and retry a compatible provider when needed.
The routes below describe the current implementation, not a permanently pinned model contract. Anarlog may update a managed route to improve quality, language coverage, latency, or availability. Select your own provider and model when you need a fixed model ID.

​
Cloud transcription

The managed Transcription route chooses a provider from the meeting’s languages and whether Anarlog needs live or after-recording transcription.
  • Soniox is the primary provider for live and after-recording transcription when it supports the selected languages. It uses Soniox 5: stt-rt-v5 live and stt-async-v5 after recording.
  • Deepgram is an eligible fallback and resolves the route to Nova 3 or Nova 2, depending on language support.
  • AssemblyAI currently resolves it to Universal 3.5 Pro: universal-3-5-pro-realtime for live transcription and universal-3-5-pro after recording.
  • The eligible fallback pool also includes configured routes for Gladia (currently Solaria 1), ElevenLabs, OpenAI, Mistral, and Alibaba Cloud Model Studio.
Only configured providers that support the requested mode and languages enter the route. If one returns a retryable error, Anarlog can try the next eligible provider. For stereo recordings, Soniox combines the microphone and computer-audio channels and uses speaker diarization to distinguish speakers.

​
Cloud Intelligence

For text tasks, Anarlog → Auto currently routes summaries, chat, and generated titles through the ~anthropic/claude-sonnet-latest alias on OpenRouter. The alias tracks the current Claude Sonnet release instead of pinning a dated model. OpenRouter receives the prompt and chooses an available endpoint for that model. Under Settings → Intelligence, Pro users can also choose Claude Sonnet, Claude Opus, GPT Sol, Gemini Pro, or Gemini Flash. Anarlog tries the selected model first, then its managed fallback route if needed. These choices track model families; use your own provider when you need a fixed model ID. Requests containing audio continue to use the audio route below. Hosted Intelligence requests that contain audio use a separate ordered model pool:
  1. google/gemini-3.1-pro-preview
  2. google/gemini-3.6-flash
  3. mistralai/voxtral-small-24b-2507
This audio-capable Intelligence route is separate from normal meeting transcription.

​
On-device models

Anarlog currently offers three built-in on-device Transcription providers. Settings shows each provider only when it is supported by your Mac and release. Parakeet Batch does not always provide word-level audio timestamps. When it returns text without a usable speech alignment, Anarlog keeps the individual words editable and groups the display by audio channel and processing chunk. The order between channels within a chunk is uncertain, and those words cannot be used for precise audio seeking. Speech-aligned words retain their timing. You can also keep audio on your network by pointing Custom, or Deepgram’s Advanced base URL, at a Deepgram-compatible server on this computer or your local network.
Compatibility code still recognizes some older local model IDs, but they are not offered as new choices. The table above is the current user-facing model list.
If you select Parakeet Streaming and also download Parakeet Batch, Anarlog can refine speaker labels on-device after a meeting with multiple remote participants. For local summaries and chat, connect LM Studio, Ollama, or Unsloth and choose any compatible model they serve. Anarlog does not silently substitute another model. Apple Intelligence is also available as an experimental, text-only Intelligence provider on eligible Macs running macOS 26 or later. If a selected local model or server is unavailable, Anarlog reports a connection error instead of sending the request to Anarlog Cloud.

​
Other model-backed features

Not every AI-adjacent feature is controlled by AI Settings:
When chat uses web search, Anarlog sends the search query to its hosted research endpoint, which currently uses Exa, even when your Transcription and Intelligence models are local. The returned text can then become context for your selected Intelligence model.

​
Your own provider

When you configure a provider with your own API key:
  • Anarlog sends requests directly from the desktop app to the configured base URL.
  • The provider and exact model shown in Settings handle the request.
  • Anarlog stores the API key in your operating system’s secure credential store, not in the app database. Keys stay on that computer and do not sync to your other devices.
  • The provider’s pricing, retention, and training policies apply.
Anarlog supports dedicated provider integrations and OpenAI-compatible endpoints. Intelligence model lists are loaded from the provider when its API supports discovery, so Settings → Intelligence is the current source of truth for available model IDs. You can also connect a subscription for Intelligence instead of an API key: Claude, ChatGPT, Grok, GitHub Copilot, or Kimi Code. Open Settings → Intelligence → Configure Providers and use Connect. Recent bring-your-own Intelligence providers include Venice, configured with a Venice API key, and Meta Muse (muse-spark-1.3 and the other muse-spark models), configured with a Meta Model API key. Recent bring-your-own Transcription models in Settings include:
  • Google Gemini 3.5 Transcribe (gemini-3.5-transcribe-live during the meeting, gemini-3.5-transcribe after recording)
  • AssemblyAI Universal 3.5 Pro (universal-3-5-pro-realtime during the meeting, universal-3-5-pro after recording)
  • Smallest AI Pulse (pulse) and Pulse Pro (pulse-pro) for live and after-recording transcription
  • Meta Muse Voice Transcribe (muse-voice-transcribe-1.0) for live and after-recording transcription with speaker labels
  • Nari Labs for live transcription with a Nari API key. Saved beta model IDs that ended in :free move to the current models.
  • Wispr Flow for transcription and dictation with your own API key. Its after-recording adapter accepts at most six minutes per request; choose another provider for longer files.
  • Inworld, Gradium, Modulate, and Alebex for live transcription with your own provider credentials.
  • NVIDIA Speech NIM for live transcription through your own deployed NIM server. NVIDIA’s hosted gRPC endpoint is not supported.
  • Amazon Bedrock for after-recording transcription through an OpenAI-compatible gateway exposing /audio/transcriptions. Native Bedrock endpoints are not supported.
The new live-only providers keep credentials unverified until transcription starts. If live transcription fails, keep the recording and retry with a provider that supports after-recording transcription. Settings may still show older model IDs for compatibility. Prefer the current Gemini 3.5, Universal 3.5, Soniox 5, Smallest AI Pulse, and GPT Transcribe models when you configure a provider yourself.

​
Choose a setup

To compare on-device, Cloud, and bring-your-own-key cost, see How can I compare the costs of all the transcription options?. For every on-device path, see How can I use on-device models?. For a fully local checklist, see Use Anarlog offline. For the data sent by each feature, see Data, privacy, and retention.