Parakeet Ultra
#7 v Přepis řeči (STT)Moondream · ultra · 3× · naposledy 24. 9. 2026
47
Momentum
Parakeet Ultra is an Open-Weight speech-to-text model published by Moondream (M87 Labs) for local transcription. It builds on NVIDIA's parakeet-tdt-0.6b-v3 and was Fine-Tuned by Moondream using the same architecture and Tokenizer to achieve lower word error rates. The model runs via Moondream's Inference Runtime "Photon" on NVIDIA GPUs and supports 25 languages as well as word and segment timestamps including proprietary voice activity detection for long audio files. Unlike its leaner sister model Parakeet Redux (ternary, 178 MB, for CPU/Apple Silicon), Parakeet Ultra is the full precision variant for GPU deployment.
Vývoj momenta
29.06.27.09.
Vlastnosti
| Real-Time Streaming | No separate streaming feature documented; model segments long audio itself via built-in VAD head (segments up to 30 seconds) |
| Latency | Runs with high real-time factor on NVIDIA GPU (e.g. B200, 128 concurrent requests in benchmarks); exact figure on official model card |
| License | CC-BY-4.0 (same as the NVIDIA base model) |
| Platform | Runs via Moondream's Photon runtime, primarily on NVIDIA GPUs (pip install moondream) |
| Price | Free, open weights (pip install moondream) |
| Release Date | September 2026 (announced on moondream.ai) |
| Languages | 25 languages (inherited from the NVIDIA base model parakeet-tdt-0.6b-v3) |