parakeet-ultra
Q4 GGUF download: handy-computer/parakeet-ultra-gguf - 0.49 GB on disk, read 2026-11-14. VRAM figures above are estimates; this is the actual file size.
Will it run on your hardware?
Pick your GPU memory - see which quantizations fit, and the cheapest card for the rest
Need an exact figure for your context length? Use the VRAM calculator.
Run it locally
Copy-paste - running in under a minute
vllm serve moondream/parakeet-ultraNew to this? Start with Ollama · serve to many users with vLLM.
Deep dive
Notes, sources, and the full write-up
parakeet-ultra is an open-weight speech recognition model published by moondream on Hugging Face in September 2026, with 630M parameters under the cc-by-4.0 license.
Among the 25 speech recognition models in this directory it ranks 25th by monthly Hugging Face downloads, and 12 of the others are larger.
HardwareLink to this section
Unquantized FP16 weights take about 1 GB; 8-bit brings that to ~1 GB and 4-bit (Q4_K_M) to ~1 GB, before activations and any context or image buffers. At this size it runs on almost any recent GPU, on Apple Silicon, and usable on a modern CPU. Check exact numbers for your card with the VRAM calculator.
Similar speech recognition modelsLink to this section
- Whisper Large V3 Turbo - 800M, 4,713,890 downloads/month
- distil-large-v3 - 760M, 745,014 downloads/month
- Qwen3-ASR-0.6B-hf - 780M, 532,640 downloads/month
- parakeet-tdt-0.6b-v3 - 630M, 531,382 downloads/month
- nemotron-3.5-asr-streaming-0.6b - 640M, 448,774 downloads/month
PopularityLink to this section
parakeet-ultra was downloaded 12,420 times on Hugging Face in the last month (70,061 all-time) and has 165 likes.
Run it locallyLink to this section
Weights and usage instructions are on the moondream/parakeet-ultra model card.
Local runtimes that fit this model:
- transformers - the reference Python library most model cards target
From the community
What users are saying
“Alright so it's working on the laptop, both cpu and cuda. For some reason I get some NeMo warnings on cuda which are actually surprising cause NeMo shouldn't be used anymore since Photon entirely replaced it for this ultra model... I need to look into it. But otherwise, working without issue.”
“Hey @moondream 🤗, it's me again, Poli from Hugging Face You keep shipping 🚢 — congrats on moondream/parakeet-ultra! Once again, me + my agent built an interactive demo app of it on Hugging Face Spaces, running on a free [ZeroGPU](https://huggingface.co/docs/hub/en/spaces-zerogpu) infrastructure.”
“Hi, impressive work! I’m baffled by how much better this model is than the original Nvidia-trained one. Thanks for sharing it so openly. Could you share some stats about the training data you used for this post-training? A per-language table with training hours would be enough.”
Frequently asked
Quick answers to common questions
How much VRAM does parakeet-ultra need?
parakeet-ultra with 630M parameters needs approximately 1 GB at Q4_K_M quantization. Use our VRAM calculator for an exact estimate.
Is parakeet-ultra better than other moondream models?
parakeet-ultra has 630M parameters - a strong choice for speech-recognition.
What license is parakeet-ultra under?
parakeet-ultra is released under the cc-by-4.0 license, making it suitable for most commercial and personal projects.
What hardware runs parakeet-ultra well?
With 630M parameters, parakeet-ultra is small enough for most consumer GPUs, Apple Silicon Macs, and many CPUs. Check our hardware directory for specific recommendations.
What is the best quantization for parakeet-ultra?
Q4_K_M is the recommended sweet spot - ~98% of FP16 quality at ~27% of the size. Q5_K_M (~1 GB) is an option if you have spare VRAM. Use our VRAM calculator to compare.
What models compete with parakeet-ultra?
parakeet-ultra competes with other models in its class. Browse our model directory for comparisons, benchmarks, and community reviews to find the best fit.
Nearby options
Similar models and compatible hardware by spec