An equal-weight average of the epoch 6, 7, 8, 9 and 10 checkpoints of No training was done here, and inference costs exactly what one model costs. Like its ingredient, this model has no reported WER or CER and cannot have one. That run trains on every held-out hour the project has, including the SAPC2 dev split the rest of this family scores against. Averaging its epochs does not create a set to measure on. So this checkpoint rests on a bet rather than a measurement, and it is worth same soup of the same five epochs was worth 0.27 CER points — 6.06% against 6.33% for the best single epoch. That is the whole of the evidence. It is evidence from a different architecture (RNN-T, not TDT) on…
Open-weight model · Speech recognition
nemotron-3.5-asr-streaming-0.6b
by NVIDIA nvidia/nemotron-3.5-asr-streaming-0.6b
/ Improve list spacing / / Badge alignment consistency / Nemotron 3.5 ASR is a multilingual, streaming Automatic Speech Recognition (ASR) model engineered to deliver high-quality multilingual transcription across both low-latency streaming and high-throughput…
Parameters638M
Context—
Weights5.7 GB
Licenseother
AccessOpen weights
Monthly Downloads748.8k
Runs On
What it takes to serve nemotron-3.5-asr-streaming-0.6b (638M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 1.3 GB | 1.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.6 GB | 0.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.3 GB | 0.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
SAVRN's Notes on nemotron-3.5-asr-streaming-0.6b
Live transcription is the job: audio in, punctuated and capitalized text out, with the chunk size set from 80 to 1,120 milliseconds. At 638 million parameters the 16-bit weights take 1.3 GB and need 1.5 GB to run, so the cheapest slot we list, one MI300X with 192 GB at $1.85 an hour, sits less than one percent occupied. That hour only pays if you stack streams on the card, and the 0.8 GB 8-bit and 0.4 GB 4-bit builds leave room for more.
The license reads other with no summary, so someone has to go through the publisher's terms before a commercial rollout. No context length is listed, so plan around chunk size. The training data names Common Voice 8.0, VoxPopuli, Europarl, FLEURS, Multilingual LibriSpeech and NVIDIA's Granary. And the weights changed on September 10, 2026, after the May 15 release, so pin the revision you validated.
Model Card
/ Improve list spacing / / Badge alignment consistency / Nemotron 3.5 ASR is a multilingual, streaming Automatic Speech Recognition (ASR) model engineered to deliver high-quality multilingual transcription across both low-latency streaming and high-throughput batch workloads. Developed by NVIDIA, this 600M parameter model transcribes speech into text with native support for punctuation and capitalization, and offers runtime flexibility with configurable chunk sizes, including 80ms, 160ms, 320ms, 560ms, and 1120ms. By leveraging a state-of-the-art Cache-Aware FastConformer-RNNT architecture, the model eliminates redundant overlapping computations common in traditional "buffered" streaming.…
Excerpt from the card by NVIDIA, licensed other.
Configuration
- Architecture
- Nemotron3_5AsrForRNNT
- Vocabulary size
- 13,088
- Model type
- nemotron3_5_asr
Identity and Version
- Repository
- nvidia/nemotron-3.5-asr-streaming-0.6b
- Publisher
- NVIDIA
- Task
- Speech recognition
- Modality
- Audio
- Library
- nemo
- Parameters
- 638M parameters
- Languages
- en, es, de, fr, it, ar, ja, ko
- Revision
- ea30d66debe3740a08b573244286791d423d6b3e
- First published
- 2026-05-15
- Last updated
- 2026-09-10
Files and Weights
22 files, 5.7 GB in total. The weights are 3 files totalling 5.7 GB in gguf, nemo, safetensors.
Weights3 files · 5.7 GB
Configuration3 files · 4.1 KB
Tokenizer2 files · 752.9 KB
Documentation5 files · 61.6 KB
Other8 files · 802.7 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 2.6 GB | 9eebdd659028 |
| nemotron-3.5-asr-streaming-0.6b.nemo | Weights | 2.4 GB | 210214ed9403 |
| nemotron-3.5-asr-streaming-0.6b.q8_0.gguf | Weights | 742.1 MB | 3fc991d3bada |
| config.json | Configuration | 1.4 KB | — |
| generation_config.json | Configuration | 193 B | — |
| processor_config.json | Configuration | 2.5 KB | — |
| README.md | Documentation | 54.2 KB | — |
| bias.md | Documentation | 2.1 KB | — |
| explainability.md | Documentation | 2.4 KB | — |
| privacy.md | Documentation | 2.1 KB | — |
| safety.md | Documentation | 737 B | — |
| arch_slide10.png | Other | 87.8 KB | 2be5d8ccfcd8 |
| avg_wer_summary.png | Other | 65.5 KB | 9a5474f81b64 |
| fleurs_langid_vs_auto.png | Other | 83.8 KB | 56c360a710c9 |
| fleurs_wer_vs_chunk_size.png | Other | 92.2 KB | 9b340101e45a |
| latency_vs_parallel.png | Other | 139.2 KB | 3e5636204786 |
| model_architecture.png | Other | 150.6 KB | 3146643d1a7c |
| model_overview.png | Other | 114.0 KB | 075554d371b3 |
| throughput_vs_chunk.png | Other | 69.5 KB | 276b2f4e3dca |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 752.1 KB | — |
| tokenizer_config.json | Tokenizer | 881 B | — |
License and Download
- License
- other
- Access
- Open weights, no gate
- Download size
- 5.7 GB
Download from NVIDIA
Released by NVIDIA through its official repository on Hugging Face.
Built From
- Described by arXiv:2305.05084
- Described by arXiv:2312.17279
- Trained on (disclosed) europarl
- Trained on (disclosed) fleurs
- Trained on (disclosed) mozilla-foundation/common_voice_8_0
- Trained on (disclosed) multilingual_librispeech
- Trained on (disclosed) nvidia/Granary
- Trained on (disclosed) voxpopuli
Evaluations
Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| ARTPARK-IISc/Vaani-Benchmark-V1.0 | Task Hindi_WERMetric Hindi_WERComparison conditions not established | 20.2 | Not named Reported by a third party |
Evaluated revision not stated | 2026-07-30 |
| FLEURS (English) | Configuration en_usTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 7.91 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (French) | Configuration fr_frTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 9.03 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (German) | Configuration de_deTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 8.31 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (Hindi) | Configuration hi_inTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 6.81 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (Italian) | Configuration it_itTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 4.25 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (Korean) | Configuration ko_krTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 7.12 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (Portuguese) | Configuration pt_brTask Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 5.48 | nvidia Publisher reported |
Evaluated revision not stated | — |
| FLEURS (Spanish) | Configuration es_419Task Automatic Speech RecognitionMetric WER (1.12s frame size, LangID)Comparison conditions not established | 4.11 | nvidia Publisher reported |
Evaluated revision not stated | — |
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 5.7 GB |
| 16-bit | 1.3 GB |
| 8-bit | 0.6 GB |
| 4-bit | 0.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Built on This Model
- Quantized fromnemotron-3.5-asr-streaming-0.6b-gguf
- Derived fromnemotron-3.5-asr-streaming-0.6b-gguf
Compare nemotron-3.5-asr-streaming-0.6b
Questions About nemotron-3.5-asr-streaming-0.6b
How much GPU memory does nemotron-3.5-asr-streaming-0.6b need?
About 1.5 GB at 16-bit and 0.4 GB at 4-bit: the weights (638M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run nemotron-3.5-asr-streaming-0.6b on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is nemotron-3.5-asr-streaming-0.6b released under?
other, as its publisher declares it. Read the license text before commercial use.
Similar Models
nvidia/parakeet-tdt-0.6b-v3 fine-tuned on all 1,047.7 hours this project holds: SAPC1 train and dev, SAPC2 train, the SAPC2 dev split the rest of this family scores against, 103.1 hours of synthetic dysarthric speech, 79.2 hours recovered by force-aligning and cutting recordings past the 45-second training cap, and 15.5 hours of AtaxiaUK and HeyJay!, which are outside the challenge corpora and make this an unconstrained-track model. This model has no reported WER or CER, and cannot have one. Every held-out hour is in its training data. That was the point: the hyperparameters were settled on the sibling runs that do hold out a dev split, and this run spends that split as training data…
This model was converted to MLX format from nvidia/parakeet-tdt-0.6b-v3 using the conversion script. Please refer to original model card for more details on the model.
This model was converted to MLX format from nvidia/parakeet-tdt-0.6b-v2 using the conversion script. Please refer to original model card for more details on the model.
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from turbo with pruned vocab, indices can be found in mapping.txt. Trained decoder only for 2^20 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. Competitive/SOTA on test sets, slightly better than 1.5B on short form, worse on long form. Also trained for…
For usage instructions follow openai/whisper-large-v3-turbo. Note for faster-whisper vocab changes make model.ismultilingual and suppresstokens wrong. Please adjust the code as required if you want to use this with faster-whisper. Turbo finetune with japanese tokenizer. Full finetune trained 2^19 steps, batch size 64. Smaller vocab with ~1.6x bytes/token allows faster speed with 4 layers vs 2 layer distil (10% larger decoder). Benchmarks. Short form slightly behind v0.2 (trained less?) but long form much better. Also trained for lyrics but untested. Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)