Lorqa/nemotron-3.5-asr-streaming-multilingual-0.6b-80ms-onnx-int8

Lorqa Nemotron 3.5 multilingual — 80 ms ONNX INT8

Byte-identical runtime mirror of csukuangfj2/sherpa-onnx-nemotron-3.5-asr-streaming-0.6b-80ms-int8-2026-06-11 at revision 2ac5952ae18a2cc010c25e3fd96ad20cf254bd09. The source model is NVIDIA Nemotron 3.5 ASR, licensed under OpenMDW 1.1. Original exporter attribution is retained in UPSTREAM_README.md.

The four runtime files total 682,215,469 bytes (about 682 MB). Encoder, decoder, joiner and tokens are unchanged. No training, re-quantization or accuracy improvements are claimed. Upstream sample recordings are not included in this mirror.

This is the multilingual model, including Mandarin Chinese; it is not the older English-only Nemotron. Chinese is a broad-coverage locale and accuracy needs evaluation on your recordings. Chunk size is 80 ms, not a guarantee of end-to-end visible text latency. Smaller chunks can trade accuracy and compute efficiency for responsiveness; running memory exceeds download size.

Runtime compatibility

Use a sherpa-onnx release supporting Nemotron 3.5 multilingual, with encoder.int8.onnx, decoder.int8.onnx, joiner.int8.onnx and tokens.txt. Per-stream language strings support explicit languages or auto. See the official multilingual deployment instructions.

This is an ONNX experiment, not a Core ML mlmodelc bundle. It cannot be loaded by Lorqa's existing Nemotron 320 ms Core ML runtime or FluidAudio Core ML loaders. App integration and actual device inference benchmarks are pending.

Mirrors: Hugging Face · ModelScope. Verify SHA256SUMS after download. PROVENANCE.json records the upstream revision and intended chunk size.