---
type: "post"
id: "parakeet-ultra-now-available-as-gguf-quantizations-for-952aab"
slug: "parakeet-ultra-now-available-as-gguf-quantizations-for-952aab"
title: "Parakeet Ultra now available as GGUF quantizations for transcribe.cpp, from F32 (2.51 GB) to Q4_K_M (486 MB)."
description: "Handy Computer has published GGUF quantizations of Moondream’s Parakeet Ultra for use with transcribe.cpp, ranging from a 2.51 GB F32 file to a 486 MB Q4_K_M file."
retrieval_nugget: "Handy Computer has published GGUF quantizations of Moondream’s Parakeet Ultra for use with transcribe.cpp, ranging from a 2.51 GB F32 file to a 486 MB Q4KM file. The release gives teams a way to test offline multilingual transcription with measured accuracy across quantizations, rather than choosing only by model size."
published_at: "2026-11-09"
updated_at: "2026-11-09"
record_date: "2026-11-09"
date_kind: "published_at"
topics: ["models","local-models","open-models"]
entities: ["Handy Computer","Parakeet Ultra","transcribe.cpp","Moondream","NVIDIA"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
source_url: "https://huggingface.co/handy-computer/parakeet-ultra-gguf"
source_urls: ["https://huggingface.co/handy-computer/parakeet-ultra-gguf"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:parakeet-ultra-now-available-as-gguf-quantizations-for-952aab"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/parakeet-ultra-now-available-as-gguf-quantizations-for-952aab/","markdown":"https://newruntime.com/posts/parakeet-ultra-now-available-as-gguf-quantizations-for-952aab.md","json":"https://newruntime.com/posts/parakeet-ultra-now-available-as-gguf-quantizations-for-952aab.json"}
markdown_url: "https://newruntime.com/posts/parakeet-ultra-now-available-as-gguf-quantizations-for-952aab.md"
---

# Parakeet Ultra now available as GGUF quantizations for transcribe.cpp, from F32 (2.51 GB) to Q4_K_M (486 MB).

## Retrieval answer

Handy Computer has published GGUF quantizations of Moondream’s Parakeet Ultra for use with transcribe.cpp, ranging from a 2.51 GB F32 file to a 486 MB Q4KM file. The release gives teams a way to test offline multilingual transcription with measured accuracy across quantizations, rather than choosing only by model size.

## Summary
Handy Computer has published GGUF quantizations of Moondream’s Parakeet Ultra for use with transcribe.cpp, ranging from a 2.51 GB F32 file to a 486 MB Q4_K_M file. The release gives teams a way to test offline multilingual transcription with measured accuracy across quantizations, rather than choosing only by model size.

## What happened
The Hugging Face model card describes the repository as a GGUF conversion of `moondream/parakeet-ultra` for transcribe.cpp. It says the conversion is pinned to upstream commit 73175eb, and reports validation against the kestrel 0.9.1 (Photon) reference at transcribe.cpp commit 08f9e6f0 on 2026-10-05. The card identifies Parakeet Ultra as Moondream’s post-trained version of NVIDIA’s `parakeet-tdt-0.6b-v3`, with 0.6 billion parameters in full precision. It is an offline speech-to-text model for the same 25 European languages as v3, not a streaming model and not a translation system.

The model takes 16 kHz mono WAV audio and produces a punctuated, cased transcript, with optional token-level timestamps. A small voice-activity head segments long recordings at pauses into chunks no longer than 30 seconds; the card says no external VAD model is needed. The release is presented as a drop-in replacement for v3, with the same size, languages and API. The repository lists six files: F32 at 2.51 GB, F16 at 1.26 GB, Q8_0 at 740 MB, Q6_K at 611 MB, Q5_K_M at 550 MB and Q4_K_M at 486 MB. It gives build instructions for transcribe.cpp and a command-line example using a WAV file; other audio must first be converted to 16 kHz mono WAV, with ffmpeg given as an example. The license is CC-BY-4.0, inherited from the base model.

The card reports a quantization comparison on the full LibriSpeech test-clean split of 2,620 utterances. At batch size 8, with no timestamps, an English language hint and CUDA decoding, F32 scores 1.80% word error rate, while Q4_K_M scores 1.82%; the other variants fall between those results. The card says the C++ F32 build matches Moondream’s kestrel runtime at 1.80% on this manifest. On 11 full TED-LIUM 3 talks, segmented by the model’s voice-activity head, F32 scores 2.22% and each quantization stays within 0.05 percentage points. The card also reports 902,523 downloads in the prior month on the repository page.

## What the archive adds
New Runtime’s earlier coverage described a Tess 4.5 27B GGUF release with files from 16.5 GB Q4_K_M to 28.6 GB Q8_0 ([2026-11-09 note](https://newruntime.com/posts/gguf-quantizations-from-q4-k-m-16-5-3b352d/)). This is a new event in the same broader distribution pattern, not a follow-up fact about Tess: Parakeet Ultra applies compact GGUF packaging to offline speech recognition, with a much smaller 486 MB low-bit option.

## Trend signal
This supports the pattern of quantized model releases lowering the hardware footprint needed to run useful workloads locally. The repository’s comparison is unusually actionable because it pairs six file sizes with reported WER results, while still depending on particular test sets and CUDA conditions. The related pattern is [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/), though this release does not itself report cost per completed task; the connection is the shift toward evaluating practical outcomes rather than model size alone.

## Why it matters
Local transcription can reduce the need to send recordings to a hosted service and can make processing viable in environments with limited connectivity or tighter data-handling requirements. A sub-500 MB quantization makes experimentation possible on more constrained machines than the 2.51 GB F32 file, and the reported gap on LibriSpeech test-clean is small. That does not establish which size is best for a production workload: teams should test their own accents, terminology, noise conditions and hardware, and compare end-to-end latency as well as recognition quality. The drop-in positioning may also make it easier to evaluate against existing transcribe.cpp use of v3, subject to verifying API compatibility in the team’s own setup.

## Limits and uncertainty
The accuracy figures are reported by the model-card authors and are not independently checked here. The LibriSpeech quantization comparison is limited to one English test split, with CUDA decoding and a stated batch size; it does not establish performance on every supported language, CPU or Apple silicon. The longer-form comparison covers 11 TED-LIUM 3 talks. The repository reports file sizes and benchmark scores but does not give a price, hosted inference option, or end-to-end latency results for the quantized builds. A model card’s claim of drop-in compatibility also needs validation in the caller’s actual application.

## Skill to build now
If your team already runs transcribe.cpp, set up a small local evaluation this week: transcribe the same representative recordings with Q4_K_M and F32, then compare WER or a human-reviewed error sample, runtime and peak memory. Include the languages, noise and domain vocabulary you actually expect. The repository’s `transcribe-cli` example accepts a 16 kHz mono WAV; convert a sample with ffmpeg first if needed. This turns the published size-versus-accuracy table into a workload-specific deployment choice.

## Reader takeaway
Parakeet Ultra is now available in transcribe.cpp-compatible GGUF files down to 486 MB, and the authors report only a small LibriSpeech test-clean WER increase for Q4_K_M versus F32. Treat that as a reason to benchmark locally, not as a universal accuracy guarantee: the published validation is bounded by its datasets and decoding setup.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)