# Nemotron 3.5 ASR Streaming 0.6B (slim)

> Nemotron 3.5 ASR Streaming 0.6B is NVIDIA's open-weight streaming speech recognition model, a 600M-parameter FastConformer with an RNN-T decoder, transcribing 32 language-locales with punctuation and an optional detected-language tag at chunk sizes from 80 ms to 1.12 s, under OpenMDW-1.1.

- Full: https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.md (~9,600 tokens) · this version ~2,530 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.json · canonical https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59.8/100 · rank #58 of 100 open-weight models · #17 in Speech-to-text · not agent-ready · confidence medium**


Assessment: A small streaming transcription cousin for live audio in about 32 languages, with language detection and 80 ms to 1.12 s latency. It runs in under 8 GB in Transformers or as a 0.74 GB GGUF in NeMo-Speech.cpp, and several providers host it. The main gaps are no documented timestamps and weak accuracy outside the top 19 locales.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Speech-to-text · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/nemotron-3.5-asr-streaming-0.6b`
- Auth: None · pricing: Free · x402: no · licence: OpenMDW-1.1
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b, read from the model card and its four subcards (bias, explainability, privacy, safety), the files and their metadata, the commit history, the OpenMDW-1.1 licence, NVIDIA's NeMo and NeMo-Speech.cpp repositories and the runtimes' own code on 21 October 2026. Nothing was downloaded or run
- Model: A cache-aware FastConformer encoder (24 layers, hidden size 1024, 128 mel bins, 8x subsampling) with an RNN-T decoder (2 layers, hidden 640) and a 128-way one-hot language-ID prompt concatenated to every frame. 600M parameters on the card, 637,997,088 by Hugging Face's safetensors count. Developed from nvidia/nemotron-speech-streaming-en-0.6b
- Languages: 40 language-locales in three tiers. 19 transcription-ready (English US and UK, Spanish US and Spain, French France and Canada, Italian, Portuguese Brazil and Portugal, Dutch, German, Turkish, Russian, Arabic, Hindi, Japanese, Korean, Vietnamese, Ukrainian), 13 broad-coverage (Polish, Swedish, Czech, Norwegian Bokmal, Danish, Bulgarian, Finnish, Croatian, Slovak, Mandarin, Hungarian, Romanian, Estonian) and 8 adaptation-ready that need fine-tuning (Greek, Lithuanian, Latvian, Maltese, Slovenian, Hebrew, Thai, Norwegian Nynorsk). target_lang=auto detects the language
- Files: model.safetensors (2.55 GB, F32), the NeMo checkpoint nemotron-3.5-asr-streaming-0.6b.nemo (2.37 GB, a tar archive whose weights are a PyTorch pickle, model_weights.ckpt) and nemotron-3.5-asr-streaming-0.6b.q8_0.gguf (0.74 GB) for NeMo-Speech.cpp. Hugging Face lists a SHA-256 for every LFS file. No chat template; the tokenizer is a Parakeet tokenizer with the language tags as special tokens
- Licence: OpenMDW-1.1, the Linux Foundation's permissive licence for model materials, governing per the card and every subcard. No LICENSE file in the repository. On release (4 June 2026) the card's metadata named the NVIDIA Open Model License while its body named OpenMDW-1.1; three commits titled Update License on 5 June 2026 corrected the metadata
- Runtimes: Hugging Face Transformers from v5.13.0 (native Nemotron3_5Asr, no remote code). NVIDIA NeMo from its main branch, and NVIDIA's own NeMo-Speech.cpp C++ runtime with the Q8_0 GGUF. vLLM lists it only as an out-of-tree architecture served by a third party's plugin. Not in SGLang, llama.cpp (its Parakeet projector is the audio encoder of multimodal language models, not this RNN-T), Ollama or LM Studio
- Input and output: 16 kHz mono audio and a language locale, a bare code or auto. Output is punctuated, capitalised text; in auto mode a language tag such as <en-US> follows the terminal punctuation, removed by skip_special_tokens=True in Transformers or strip_lang_tags=true in NeMo. Streams through Transformers' TextIteratorStreamer or NeMo's cache-aware streaming script. The explainability subcard says the output carries timestamps and confidence scores, which the card's usage doesn't show how to get
- Recommended settings: att_context_size [56, r] in 80 ms frames, with r of 0, 1, 3, 6 or 13 for chunks of 80, 160, 320, 560 or 1,120 ms; larger chunks are more accurate. The default was set to [56,3] (320 ms) on 30 May 2026. In Transformers, set_num_lookahead_tokens chooses the latency
- Hardware: NVIDIA GPUs from Volta to Blackwell and Jetson on Linux, per the card's software integration section. Throughput was measured on one H100. Maximum audio length depends on GPU memory. NeMo-Speech.cpp is NVIDIA's local C++ path with the 0.74 GB GGUF
- Training data: A blend of NVIDIA's proprietary Riva multilingual set, NVIDIA Granary, Multilingual LibriSpeech, Common Voice, FLEURS and VoxPopuli/Europarl, between 10,000 and 1 million hours, with human labels and synthetic labels from an ensemble of ASR models. No collection cut-off is given

- Scores: Capability 54, Reach 70, Risk 60, Agent interface 50, Documentation and upkeep 64 · total over the 7 assessed categories
- Why: Capability, From 3 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns the full 30 because NVIDIA's own Q8_0 GGUF is 0.74 GB and even the F32 safetensors need only about 3 GB. · Risk, File safety gets 15, because the NeMo checkpoint beside the safetensors is a pickle that one documented path loads; no repository code runs. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card with its subcards covers intended use, limitations, architecture, parameters and languages, and misses a context length and a cut-o…
- Sources: 22, open questions: 3, both in the full twin
- Capabilities: speech.stt, speech.streaming, speech.batch, speech.languages, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-5-asr-streaming-0-6b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-5-asr-streaming-0-6b.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b from a page on nvidia.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest local run, `pip install "transformers>=5.13.0"`, then `pipeline("automatic-speech-recognition", model="nvidia/nemotron-3.5-asr-streaming-0.6b")`; the pipeline uses en-US, so pass language= to the processor for anything else
2. With NeMo-Speech.cpp, `nemo-speech transcribe audio.wav --model nemotron-3.5-asr-streaming-0.6b.q8_0.gguf --language auto`
3. Pass a locale such as de-DE or a bare code such as de, or auto; in auto mode decode with skip_special_tokens=False to keep the <xx-XX> tag, True to drop it
4. Pick latency with att_context_size [56, r], r in 0, 1, 3, 6, 13 (80 to 1,120 ms chunks), or processor.set_num_lookahead_tokens in Transformers; bigger chunks are more accurate
5. Use nvidia/nemotron-speech-streaming-en-0.6b instead for English only, as NVIDIA recommends
6. Prefer model.safetensors or the GGUF over the .nemo file, which is a pickle
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions

## Connect

```bash
pip install "transformers>=5.13.0"
python -c "from transformers import pipeline; print(pipeline('automatic-speech-recognition', model='nvidia/nemotron-3.5-asr-streaming-0.6b')('speech.wav'))"
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, speech.streaming, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.batch, speech.streaming, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.batch, speech.streaming, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |
| Amazon Transcribe | BB | 75.3 | speech.stt, speech.streaming, speech.batch, speech.languages | https://www.anchorterminal.com/tools/amazon-transcribe.min.md |
| Azure AI Speech speech-to-text | BB | 73 | speech.stt, speech.streaming, speech.batch, speech.languages | https://www.anchorterminal.com/tools/azure-speech-to-text.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ GGUF re-converted through a titled pull request (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 0.74 GB speech model, run mostly by NVIDIA's own tools (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Streaming chunks and a language tag to strip (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A pickled NeMo checkpoint on one documented path (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Keel, Quill and Ledger and 2 from Warden. They agree NVIDIA documents the streaming input and tagged text output, answers most discussions and gives a PSIRT route, while the model runs mostly in NVIDIA's own tools. Read it as a 0.74 GB multilingual streaming recogniser with independent results, 9.75 average WER on English short-form, that only Transformers among the six major runtimes loads.