What is Nemotron 3.5 ASR Streaming 0.6B?
Nemotron 3.5 ASR is a 600-million-parameter streaming speech recognition model from NVIDIA, posted on Hugging Face on 4 June 2026 under the permissive OpenMDW-1.1 licence. It is the multilingual extension of NVIDIA's English Nemotron Speech Streaming model, adding a language prompt so one checkpoint covers 40 language-locales: 19 NVIDIA calls transcription-ready, 13 broad-coverage and 8 that need fine-tuning before use.
It uses a cache-aware FastConformer encoder with an RNN-T decoder, which keeps the encoder state between chunks instead of re-processing overlapping audio. Chunk size, and so latency, is chosen at run time from 80, 160, 320, 560 or 1,120 milliseconds without retraining. Output includes punctuation and capitals, and with the language set to auto it appends a tag naming the language it heard. It runs in NeMo, Hugging Face Transformers 5.13.0 or later, and NVIDIA's C++ NeMo-Speech.cpp runtime from a q8_0 GGUF file.
Is Nemotron 3.5 ASR Streaming 0.6B worth it?
Nemotron 3.5 ASR is a capacity tool for live transcription. On one H100 NVIDIA measures it sustaining 240 concurrent real-time streams at the 80 ms setting, against 14 for its older Parakeet RNNT 1.1B, and 2,400 streams against 400 at 1,120 ms. For a voice-agent or captioning platform, that is the difference between one GPU and a rack, and the latency knob lets you trade delay for accuracy per deployment rather than per model.
Accuracy is middling and NVIDIA's tables say so. On FLEURS with the language supplied and 1.12-second chunks it scores 4.11% WER in Spanish, 7.91% in English, 8.31% in German and 9.03% in French, averaging 8.84% across its 19 transcription-ready locales; at 80 ms that average rises to 10.38%. Offline models do far better: NVIDIA's own Nemotron 3.5 Transcribe sits at 3.4% AA-WER on Artificial Analysis, but it cannot stream. NVIDIA itself recommends the English-only streaming model for English work.
The broad-coverage tier is where buyers should read carefully. Polish, Mandarin and Czech come in at 15.15% to 20.41% error and Danish and Hungarian at 27.49% and 28.68%, averaging 22.13%, which is roughly one word in five wrong. The 8 adaptation-ready locales, including Greek, Hebrew and Thai, produce no usable transcripts until you fine-tune them; NVIDIA publishes a fine-tuning guide for that.
The catch: this is NVIDIA-shaped software. The supported platforms are Linux and Jetson on NVIDIA GPUs, the input must be mono, and part of the training labels were generated by an ensemble of other models including Whisper, so systematic errors those models share can carry over. There are no speaker labels.
Our take: Nemotron 3.5 ASR is the engine to put under live captions in Spanish, Italian, Portuguese or Hindi at scale; for recordings that can wait a few seconds, an offline model will make noticeably fewer mistakes.
Where it earns its place
A platform engineer at a contact-centre software company
Give live agents a running transcript of every call in Spanish, Portuguese and English
Serve the model with NeMo's cache-aware streaming script at 560 ms or 1,120 ms chunks; NVIDIA measured 2,400 parallel streams per H100 at 1,120 ms. Hosted on DeepInfra, a million minutes a month is about $200.
A developer building a voice agent for an Indian bank
Recognise Hindi and English callers fast enough for natural turn-taking
Run at 80 ms or 160 ms with target_lang=auto so each utterance carries a language tag; Hindi scores 8.13% WER at 80 ms with the language supplied, and 11.47% when auto-detected.
An ML team at a Greek broadcaster
Caption live news in Greek
Greek is in the adaptation-ready tier, so budget for fine-tuning on in-house transcripts using NVIDIA's guide before going live; out of the box the model will not transcribe it reliably.
Before you commit
- Setup
- pip install NeMo from GitHub (Python 3.11+, libsndfile and ffmpeg), load nvidia/nemotron-3.5-asr-streaming-0.6b with from_pretrained and run the cache-aware streaming script with att_context_size and target_lang; or transformers 5.13.0 pipeline.
- Integrations
- NeMo, Hugging Face Transformers (offline and streaming), NeMo-Speech.cpp with a q8_0 GGUF, NVIDIA Speech NIM; DeepInfra via OpenRouter for a hosted endpoint.
- Data and privacy
- OpenMDW-1.1 permits commercial use. Self-hosted, audio stays on your hardware. Training mixed public sets, NVIDIA's proprietary Riva data and labels generated by other ASR models.
- Cost at scale
- Self-hosted, one H100 handled 2,400 streams at 1,120 ms latency in NVIDIA's test. Hosted on DeepInfra at $0.00000333 a second, a million minutes is about $200.
- Limits
- 32 locales usable out of the box, 8 more after fine-tuning; mono input; Linux and NVIDIA GPUs; no diarization; accuracy drops as chunk size shrinks.
What’s good
- 240 streams at 80 ms and 2,400 at 1,120 ms on one H100 in NVIDIA's tests
- Latency chosen at run time, from 80 to 1,120 ms
- 40 language-locales in one model with automatic language tagging
- Permissive OpenMDW-1.1 licence; GGUF build for NeMo-Speech.cpp
- About $0.20 per 1,000 minutes hosted
What’s not
- 7.91% English WER on FLEURS at its slowest setting
- Broad-coverage languages average 22.13% WER
- 8 locales need fine-tuning before they work
- Linux and NVIDIA GPUs only; mono audio
- No speaker labels
Best for
Platforms running many simultaneous live transcription or voice-agent sessions on NVIDIA GPUs.
Nemotron 3.5 ASR Streaming 0.6B alternatives
Nemotron 3.5 ASR Streaming 0.6B: common questions
- Is Nemotron 3.5 ASR Streaming 0.6B worth it?
- Nemotron 3.5 ASR Streaming is built for running hundreds of live captions on one GPU, and its accuracy is the price of that density. We rate it 3.9 out of 5.
- How much does Nemotron 3.5 ASR Streaming 0.6B cost?
- Free under OpenMDW-1.1. Hosted by DeepInfra on OpenRouter at $0.00000333 a second of audio, about $0.20 per 1,000 minutes.
- What are the downsides of Nemotron 3.5 ASR Streaming 0.6B?
- 7.91% English WER on FLEURS at its slowest setting; broad-coverage languages average 22.13% WER; 8 locales need fine-tuning before they work; linux and NVIDIA GPUs only; mono audio; no speaker labels.
- Who is Nemotron 3.5 ASR Streaming 0.6B best for?
- Platforms running many simultaneous live transcription or voice-agent sessions on NVIDIA GPUs.
- How much does Nemotron 3.5 ASR Streaming 0.6B cost for a team or in production?
- Self-hosted, one H100 handled 2,400 streams at 1,120 ms latency in NVIDIA's test. Hosted on DeepInfra at $0.00000333 a second, a million minutes is about $200.
- Is Nemotron 3.5 ASR Streaming 0.6B safe for company data?
- OpenMDW-1.1 permits commercial use. Self-hosted, audio stays on your hardware. Training mixed public sets, NVIDIA's proprietary Riva data and labels generated by other ASR models.
- What are the best Nemotron 3.5 ASR Streaming 0.6B alternatives?
- The closest alternatives we have reviewed are NVIDIA Nemotron 3.5 Transcribe, MAI-Transcribe-2-Streaming, Qwen3-ASR (1.7B and 0.6B), Parakeet TDT 0.6B v3 and Voxtral 3 (Transcribe and TTS).
Answers checked Nov 7, 2026.