Speech models

Hyprcore ships more than a dozen speech-to-text models across seven families, and every one of them runs locally on your Mac with no internet required. You can also use cloud providers when you want speed or live streaming. You set the model in two places—Settings → Models → Dictation and Settings → Models → Meetings—because dictation and meetings often want different tradeoffs.

Local model families

Family	Models	Best for	Notes
Whisper	Small, Medium, Turbo, Large	Meetings and general use	The all-rounder. 99 languages. Small/Medium are fast; Large is most accurate; Turbo scales best on long audio. Apple Silicon recommended for Turbo and Large.
Parakeet	V2, V3	Dictation only	CPU-only, ~5× real-time. V3 auto-detects 25 European languages and is the recommended default for dictation; V2 is English-only. Not available for meetings.
NVIDIA Canary	180M Flash, 1B Flash, 1B v2	Multilingual meetings	NVIDIA's NeMo Canary models, with translation. Flash variants cover English, German, Spanish, and French; 1B v2 covers 25 European languages at the highest accuracy.
Moonshine	Base, V2 Tiny, V2 Small, V2 Medium	Fast English dictation	Very fast, English-only. Dictation only.
SenseVoice	SenseVoice	Chinese, Japanese, Korean, Cantonese	Very fast Asian-language recognition. Meeting-capable.
Breeze	Breeze ASR	Taiwanese Mandarin	Optimized for Taiwanese Mandarin with code-switching. Runs on the Whisper engine, so it works for meetings too.
GigaAM	GigaAM v3	Russian	Fast, accurate Russian recognition. Dictation only.

All models are quantized for size and speed without obvious quality loss. Sizes on disk range from about 31 MB (Moonshine V2 Tiny) to roughly 1.6 GB (Whisper Turbo).

Note: Parakeet, Moonshine, and GigaAM are dictation-only—they don't handle long meeting audio safely, so they won't appear in the meetings model picker. Whisper (including Breeze), Canary, and SenseVoice are meeting-capable.

Which one should I use

• Just dictating? Parakeet V3. Auto-language, fastest CPU performance, no GPU dependency.

• Recording short meetings (under 30 minutes)? Whisper Medium. Good summary inputs.

• Recording long meetings or podcasts? Whisper Turbo on Apple Silicon. It scales better with hours of audio.

• Multilingual meetings? NVIDIA Canary. The Flash variants cover English, Spanish, French, and German; Canary 1B v2 stretches to 25 European languages.

• Need maximum accuracy and you have an M-series Mac? Whisper Large.

Cloud models

Cloud STT runs on a provider's servers. Faster, supports live streaming, and lets you use diarization (who said what). The tradeoff is that audio leaves your Mac, and you pay for usage.

Provider	Modes	Diarization	Cost
Deepgram	Batch + live	Yes (best in class)	Hyprcore Cloud tokens or BYOK
Groq Whisper	Batch	Limited	Hyprcore Cloud tokens or BYOK
OpenAI Whisper API	Batch	No	BYOK only

To use cloud STT, configure the provider in Settings → Models → Speech provider. Pick Hyprcore Cloud if you want to pay with your monthly tokens and not manage API keys, or Bring your own key to use your own Deepgram/Groq/OpenAI account.

Provider policy

In Settings → Meetings → STT provider policy, choose how Hyprcore picks an engine:

• Prefer local (default). Use the local model unless you explicitly switch.

• Prefer cloud. Use the cloud provider when available; fall back to local on errors.

• Local only. Never call the cloud, even if a provider is configured.

• Cloud only. Always use the cloud; fail if it's not available.

• Manual. Always ask which engine to use when starting a meeting.

Downloading

When you select a model that's not yet on disk, Hyprcore downloads it from models.hyprcore.ai. Download progress shows in the model picker. Models are cached at:

~/Library/Application Support/ai.hyprcore.desktop/models/

You can delete a model file to free space; Hyprcore re-downloads on next use.

If you're on a restricted network, see Manual model install.