Skip to content
Audio

Nemotron 3.5 ASR Streaming 0.6B (CoreML)

fast speech recognition for Apple Silicon

2 min readupdated

A ready-to-use build of NVIDIA's streaming automatic speech recognition (ASR) model, converted to CoreML to run locally on Apple Silicon via the Apple Neural Engine (ANE — Apple's built-in neural processor). Based on a 0.6B-parameter Conformer + RNN-T architecture, this version skips heavy server setups—it uses post-training quantization and comes pre-compiled to .mlmodelc so you can drop it straight into iOS 17+ and macOS 14+ apps. Built for iOS and macOS developers who need real-time offline transcription.

Key features

  • Two vocabulary options: a lightweight latin bundle (2,828 tokens covering en, es, fr, it, pt, de) and a full multilingual bundle (13,087 tokens supporting Chinese, Japanese, and over 100 other languages via prompt_id).
  • Four latency modes: pre-packaged bundles for 0.56s, 1s, 2s, and 4s chunks let you trade off responsiveness against throughput (2s is recommended by default).
  • Lossless mixed-precision quantization: outer encoder layers are compressed to INT8 and middle layers to 6-bit, cutting encoder size by roughly 55% compared to FP16 without hurting accuracy.
  • M2 and M3 fixes: the repo includes alternate encoder variants (encoder-v2 and a split encoder-split) that compile cleanly on M2 and M3 Neural Engines without falling back to the CPU.

What you can do

  • Offline captions and voice recorders on iPhone and Mac: embed the compiled bundles into mobile or desktop apps via the FluidAudio library for instant transcription without an internet connection.
  • Voice assistants and local agents: use the 1s tier (for European languages) or 2s tier (for multilingual setups) as a fast speech-input front end on a MacBook.
  • Batch transcription of long recordings: plug in the 4s bundle for high-speed offline transcription of lectures and interviews.

Example

fluidaudiocli nemotron-multilingual-transcribe \
    --input audio.wav \
    --model-dir latin/2240ms \
    --language de-DE
Read moreDetails
• Speed on Apple M5 Pro (2s chunk): 130–150 RTFx for Latin languages (4.80%–9.83% WER on FLEURS) and 80–89 RTFx for the multilingual version; on LibriSpeech test-clean for English, it hits 120 RTFx at 3.67% WER.
• The full-vocabulary version (multilingual) is sensitive to chunk size: on 0.56s and 1s chunks, speed for CJK drops to 21–27 RTFx, so stick to the 2s mode for those languages.
• Compatibility: iOS 17+ / macOS 14+, distributed under the commercial OpenMDW-1.1 license (repo access is granted upon request in the author's Discord).
Discuss Telegram 0 21

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.5 GB · ready files1.1 GB
  • FP8 / INT8recommendedweights 0.3 GB0.8 GB
  • 4-bitweights 0.1 GB0.7 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # mustafagunes/nemotron-coreml
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download mustafagunes/nemotron-coreml --local-dir ~/models/nemotron-coreml
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to nemotron-coreml
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of nemotron-3.5-asr-streaming-0.6b (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face
Merges 1All on Hugging Face

About the model

mustafagunes/nemotron-coreml
Author
mustafagunes
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments