---
type: "post"
id: "upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762"
slug: "upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762"
title: "Upstream Ultra cuts FLEURS 25-language WER from 11.62% to 9.55% vs base parakeet-tdt-0.6b-v3."
description: "Moondream’s Parakeet Ultra is a post-trained version of NVIDIA’s 0.6-billion-parameter Parakeet TDT 0.6B v3, with lower reported word error rates across the benchmark groups on its model card."
retrieval_nugget: "Moondream’s Parakeet Ultra is a post-trained version of NVIDIA’s 0.6-billion-parameter Parakeet TDT 0.6B v3, with lower reported word error rates across the benchmark groups on its model card. The strongest headline is FLEURS across 25 languages: average WER falls from 11.62% to 9.55%. For builders, it is a candidate for testing multilingual transcription without a larger model."
published_at: "2026-11-09"
updated_at: "2026-11-09"
record_date: "2026-11-09"
date_kind: "published_at"
topics: ["models","evals","open-models"]
entities: ["Moondream","Parakeet Ultra","Parakeet TDT 0.6B v3","NVIDIA","Photon","FLEURS"]
related_patterns: ["ai-engineering-evaluations-are-splitting-into-task-specific"]
novelty_label: "new_fact"
source_url: "https://huggingface.co/moondream/parakeet-ultra"
source_urls: ["https://huggingface.co/moondream/parakeet-ultra"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762"
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/upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762/","markdown":"https://newruntime.com/posts/upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762.md","json":"https://newruntime.com/posts/upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762.json"}
markdown_url: "https://newruntime.com/posts/upstream-ultra-cuts-fleurs-25-language-wer-from-9eb762.md"
---

# Upstream Ultra cuts FLEURS 25-language WER from 11.62% to 9.55% vs base parakeet-tdt-0.6b-v3.

## Retrieval answer

Moondream’s Parakeet Ultra is a post-trained version of NVIDIA’s 0.6-billion-parameter Parakeet TDT 0.6B v3, with lower reported word error rates across the benchmark groups on its model card. The strongest headline is FLEURS across 25 languages: average WER falls from 11.62% to 9.55%. For builders, it is a candidate for testing multilingual transcription without a larger model.

## Summary
Moondream’s Parakeet Ultra is a post-trained version of NVIDIA’s 0.6-billion-parameter Parakeet TDT 0.6B v3, with lower reported word error rates across the benchmark groups on its model card. The strongest headline is FLEURS across 25 languages: average WER falls from 11.62% to 9.55%. For builders, it is a candidate for testing multilingual transcription without a larger model.

## What happened
Moondream’s Hugging Face model card describes Parakeet Ultra as a post-trained version of `parakeet-tdt-0.6b-v3`. The card says it retains the original architecture, tokenizer and 0.6-billion-parameter size in full precision. It reports that Ultra performs better than the original across each listed benchmark group and runs faster than the original does in NeMo. The model’s license is CC-BY-4.0, the same as the original’s.

On the Hugging Face Open ASR Leaderboard’s seven English test sets, the card gives an average WER of 5.80% for Ultra versus 6.26% for the original. The listed sets cover audiobooks, meetings, earnings calls, podcasts and YouTube, financial calls, and parliamentary speech. In the FLEURS test split, which covers read Wikipedia sentences in all 25 supported languages, the average is 9.55% versus 11.62%. Ultra’s reported WER is lower in every listed language, including Russian (5.21% versus 5.91%) and Ukrainian (4.75% versus 5.94%). The language-level figures vary considerably: the card reports 2.00% for Italian and 32.25% for Greek with Ultra, illustrating why the aggregate alone cannot predict performance on a particular language.

The card also reports results for business speech, background noise and long-form audio. Its business speech average is 5.79% versus 6.15% for the original, using an AA-WER-style setup: cleaned AMI and VoxPopuli test sets plus Earnings-22 scored in 30-second chunks joined per call. For nine MUSAN noise conditions, the reported average is 5.82% versus 6.72%; the conditions include LibriSpeech and German and Spanish FLEURS audio with added noise at 10, 5 and 0 dB signal-to-noise ratios. On eleven complete TED-LIUM 3 talks, each 10–20 minutes long, WER is 1.94% versus 2.71%. The card says Photon uses the model’s voice-activity head to split long recordings at pauses into segments of at most 30 seconds, without an external VAD model.

The prescribed usage is through Photon: install the `moondream` Python package and load `moondream/parakeet-ultra`. Its transcription interface can return text and, optionally, sentence- or word-level timestamps. Moondream says every accuracy figure on the page was produced with Photon on an NVIDIA GPU. The card says scores were calculated with the Open ASR Leaderboard’s own pipeline as of September 2026, including its normalizers and compound-merging alignment. Its speed comparison used one NVIDIA B200 and 128 requests in flight; NeMo 3.0 ran the original checkpoint at batch 128 while Photon ran Ultra. The page reports higher real-time factors for Ultra on LibriSpeech test-clean and AMI. The card says the model is not deployed by an Inference Provider on Hugging Face.

## What the archive adds
We have not covered this before; it opens a new thread.

## Trend signal
This does not yet fit a pattern New Runtime tracks. The model card reports benchmark improvements, but it does not establish a broader shift in how transcription systems are built or deployed.

## Why it matters
For teams processing multilingual audio, an improvement on a fixed-size model is worth testing: it may offer lower error rates without requiring a larger parameter count. The Russian and Ukrainian FLEURS results are relevant signals for teams whose corpora include those languages, while the long-form and noise tests point to other conditions to reproduce in evaluation. Keeping the same architecture and tokenizer also makes this a more focused comparison than a move to a larger model family, although post-training remains a change that should be checked against a team’s own audio.

The comparisons do not settle which model is better for a particular workflow. Builders should run both candidates against representative recordings, compare normalized transcripts and timestamps, and include the cost and throughput of their own inference setup. Photon and an NVIDIA GPU are part of the reported setup, so results may not transfer unchanged to another runtime or hardware. In particular, teams should not assume that aggregate FLEURS gains mean equal improvement across every language, acoustic condition or vocabulary.

## Limits and uncertainty
The reported figures come from Moondream’s model card, not an independent evaluation described here. The card specifies the benchmark sets and evaluation conditions, but those test results do not establish accuracy on every language variety, accent, recording environment or domain. The FLEURS split uses read Wikipedia sentences, not conversational audio, and the business and noise evaluations cover defined test sets rather than all real production conditions. The speed comparison uses a B200, 128 concurrent requests, and different runtimes for the two models (Photon for Ultra and NeMo for the original), so it is not a controlled comparison of model weights alone. The card does not provide a hosted inference option through Hugging Face Inference Providers.

## Skill to build now
Take a representative sample of your team’s recordings and score both Parakeet Ultra and your current ASR system against corrected transcripts. Include the languages, noise levels, accents and recording lengths that matter in production; test word-level timestamps if downstream alignment depends on them. Keep runtime, hardware and concurrency identical where possible, and measure both WER and throughput before switching. Report results by language and condition rather than relying only on one average, and include failures that make transcripts unusable even when their word error rate looks acceptable.

## Reader takeaway
Parakeet Ultra reports lower WER than its 0.6B-parameter predecessor across several benchmark groups, including FLEURS in 25 languages. Treat it as a strong candidate for a local evaluation, not as proof that every multilingual transcription workload will improve. The useful next step is a side-by-side test on your own recordings and deployment 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: [Evals](https://newruntime.com/topics/evals.md): Explore the evals 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)