Parakeet Ultra arrives on OpenRouter

The meaningful update is distribution: a small, accurate transcription model is easier to test and integrate without managing GPU inference.

• –Keeps the same architecture and parameter count as NVIDIA’s Parakeet TDT 0.6B v3 while improving reported word-error rates.
• –Moondream reports 5.80% English WER versus 6.26% for the original, plus stronger results on multilingual, noisy, and long-form audio.
• –Segment and word timestamps make it useful for captions, meeting notes, voice agents, and searchable audio.
• –Ultra is optimized for GPUs, while Moondream’s Parakeet Redux targets CPUs and Apple silicon, giving developers a clearer deployment tradeoff.
• –Hosted access lowers experimentation friction, but teams still need to evaluate latency, provider reliability, and data-handling requirements.