Lorqa Nemotron 3.5 multilingual — 160 ms ONNX INT8
Byte-identical runtime mirror of csukuangfj2/sherpa-onnx-nemotron-3.5-asr-streaming-0.6b-160ms-int8-2026-06-11 at revision b3a4dbde84fba1a13cb4270e6730b525ac6a2db6. 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 682,215,471 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 160 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.