Horch vs Kyutai Moshi

A side-by-side look at pricing, capabilities, pros, cons, and our editorial scores.

	Horch Audio	Kyutai Moshi Audio
Tagline	Privacy-first, on-device meeting assistant for macOS	Open-source, full-duplex speech-to-speech foundation model with sub-200ms latency
Category	Audio	Audio
Pricing	Paid· One-time purchase: €49 once	Free· Free and open source. Models under CC-BY 4.0, code under MIT (Python) / Apache 2.0 (Rust). Self-hosted only — you pay your own compute (24GB+ GPU for PyTorch, or Apple Silicon via MLX).
Model	Whisper (local) for transcription; optional Ollama / MLX local LLMs for summarization	Moshi (7B-class speech-text foundation model) + Mimi neural audio codec, in-house by Kyutai
Editorial score	—	—
Use cases	Confidential client meeting transcriptionAutomatic action-item extractionPer-contact relationship historyPre-meeting briefings from prior callsLocal Whisper transcription without cloud uploadFeeding meeting context into Claude or Cursor via MCPPersonal second-brain in MarkdownLegal, medical, or NDA-bound conversation notes	Real-time voice assistant prototypesResearch on full-duplex spoken dialogueOn-device voice interaction on Apple Silicon via MLXLow-latency conversational agents behind WebSocketNeural audio codec experimentation with MimiSelf-hosted voice interface for privacy-sensitive appsSpeech tokenizer for downstream audio LLM trainingInterruptible in-car or wearable voice UX
Pros	Fully on-device by default — audio and transcripts never leave the Mac unless the user opts in One-time €49 purchase instead of a recurring per-seat SaaS bill Records at OS level so no bot appears in the meeting and any app (Zoom, Meet, Teams, in-person) works Notes stored as plain Markdown in ~/Meetings — portable, greppable, Obsidian-friendly Auto-generates action items, topic summaries, and per-person profiles from spoken content MCP server exposes meeting history to Claude, Cursor, and other agent clients Local Whisper transcription plus optional Ollama/MLX means you pick the summarizer	Truly full-duplex — handles interruptions, overlap and back-channels rather than rigid turn-taking Sub-200ms practical latency on a single L4 GPU, well below third-party voice APIs Fully open weights (CC-BY 4.0) plus MIT/Apache code — self-host with no per-minute billing Ships with Mimi, a streaming neural audio codec that beats SpeechTokenizer and SemantiCodec Multiple inference backends: PyTorch for research, Rust/Candle for production, MLX for on-device Mac/iPhone Inner-monologue text prediction gives you a transcript alongside the audio stream for free
Cons	macOS only — no Windows, Linux, iOS, or web client Local Whisper and summarization need a reasonably modern Apple Silicon Mac to feel fast No cloud sync or team workspace, so sharing across a team requires bring-your-own storage Small independent product without the integrations catalog of Fireflies, Otter, or Fathom OS-level capture depends on macOS screen/audio permissions, which some corporate MDM setups block	English-only voices at launch — no multilingual support out of the box Knowledge and reasoning quality trail top text LLMs; it's a 7B-class model, not GPT-4o Voice Requires a 24GB+ GPU for the reference PyTorch build; on-device is only viable via MLX on Apple Silicon No hosted API or SaaS tier — you own the ops, scaling and safety filtering Only two fixed synthetic voices (Moshiko/Moshika); no voice cloning or speaker conditioning in the release
Website	horch.app	kyutai.org

Pick Horch if

✅ Fully on-device by default — audio and transcripts never leave the Mac unless the user opts in
✅ One-time €49 purchase instead of a recurring per-seat SaaS bill
✅ Records at OS level so no bot appears in the meeting and any app (Zoom, Meet, Teams, in-person) works
✅ Notes stored as plain Markdown in ~/Meetings — portable, greppable, Obsidian-friendly

Pick Kyutai Moshi if

✅ Truly full-duplex — handles interruptions, overlap and back-channels rather than rigid turn-taking
✅ Sub-200ms practical latency on a single L4 GPU, well below third-party voice APIs
✅ Fully open weights (CC-BY 4.0) plus MIT/Apache code — self-host with no per-minute billing
✅ Ships with Mimi, a streaming neural audio codec that beats SpeechTokenizer and SemantiCodec

Compare a different pair →