Is Moondream Parakeet Ultra state of the art?Pending benchmarks.Announced Sep 22, 2026; no independent results yet.
Moondream's post-trained version of NVIDIA's Parakeet TDT 0.6B v3 speech-to-text model, with the same architecture, tokenizer and 0.6B parameters in full precision. It transcribes 25 European languages with segment and word timestamps, and its weights carry a small voice-activity head that Moondream's Photon engine uses to split long audio at pauses. Open weights on Hugging Face under CC-BY-4.0; on OpenRouter since Nov 3. Not an LLM.
The details so far
- Lab
- Moondream
- Announced
- Sep 22, 2026
- API
- Live · moondream/parakeet-ultra
- Open weights
- Published · Hugging Face (moondream/parakeet-ultra), FP32 safetensors
- Parameters
- 0.6B
- Task
- Speech recognition (speech to text)
- Base model
- NVIDIA parakeet-tdt-0.6b-v3, post-trained; same architecture, tokenizer and output conventions
- Languages
- 25: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovene, Spanish, Swedish, Ukrainian
- Timestamps
- Segment and word level (Photon transcribe API)
- Long audio
- Built-in voice-activity head; Photon cuts recordings at pauses into segments of up to 30 seconds
- License
- CC-BY-4.0, same as the original
- Runtimes
- Photon (Moondream); drop-in for Parakeet engines; parakeet.cpp support announced Oct 4
- API price
- $0.06 per hour of audio on OpenRouter (Prime Intellect), listed Nov 3
- Sibling
- Moondream Parakeet Redux, a ternary 178 MB version for CPUs, released the same week (not tracked here)
- Scores
- Moondream's word error rates are self-run, so they add no rows to the Speech-to-text board
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: lower word error rate than parakeet-tdt-0.6b-v3 in every benchmark group Moondream tested: Open ASR Leaderboard English sets 5.80% vs 6.26%, FLEURS 25 languages 9.55% vs 11.62%, business speech 5.79% vs 6.15%, background noise 5.82% vs 6.72%, TED-LIUM long-form 1.94% vs 2.71%.
Lab claim: on one NVIDIA B200, Photon runs it at 9,743× real time on LibriSpeech test-clean, against 6,005× for the original in NeMo 3.0.
Sources
LABIntroducing Moondream Parakeet Redux and Parakeet Ultra (Moondream blog)moondream.ai/blog/introducing-parakeet-redux-and-ultra · Sep 22, 2026LABMoondream co-founder vikhyatk release post on Xx.com/vikhyatk/status/2102372706658447712 · Sep 22, 2026LABMoondream release post on Xx.com/moondreamai/status/2102494472106119401 · Sep 22, 2026REPOmoondream/parakeet-ultra on Hugging Facehuggingface.co/moondream/parakeet-ultraLABparakeet.cpp support post on X (vikhyatk)x.com/vikhyatk/status/2106652266929525017 · Oct 4, 2026LABOpenRouter availability post on X (vikhyatk)x.com/vikhyatk/status/2117735669539954876 · Nov 3, 2026PRESSParakeet Ultra on OpenRouteropenrouter.ai/moondream/parakeet-ultra · Nov 3, 2026Tracking continues until an independent board lists it.