Modelpublic

GoodnotesIT/nemotron-3.5-asr-streaming-multilingual-0.6b-litert-fp16

sourceHugging Faceotherupdated 11d agoView on Hugging Face

0likes189downloads

Model Card

Nemotron 3.5 ASR Streaming Multilingual 0.6B — LiteRT FP16

Goodnotes-owned LiteRT export of `nvidia/nemotron-3.5-asr-streaming-0.6b` from the checksum-pinned NVIDIA .nemo checkpoint.

Included artifacts

—portable/: FP16 weights with FP32 I/O for on-device NPU, GPU, then CPU JIT fallback.
—aot/<vendor>/<SoC>/encoder.tflite: AOT encoder candidates for mediatek/MT6878, mediatek/MT6897, mediatek/MT6983, mediatek/MT6985, mediatek/MT6989, mediatek/MT6991, qualcomm/SM8450, qualcomm/SM8550, qualcomm/SM8650, qualcomm/SM8750, qualcomm/SM8850, samsung/E9965.
—reports/: complete cross-compiler result, including unsuccessful vendor targets.
—release-manifest.json: byte lengths, SHA-256 checksums, source, and toolchain identity.

Only the encoder is packaged per AOT target because it is the graph executed on the selected accelerator. Decoder and joint remain shared portable graphs. Incomplete targets are excluded and use the portable JIT chain.

Artifact revision: 03351ad96f2789192d558b7005764e58bc06f766 LiteRT exporter/compiler: 2.3.0.dev20260913 Android compatibility interpreter: 2.1.6

The portable graphs have executed successfully with representative typed inputs through the LiteRT 2.1.6 CPU interpreter. The AOT files are compiler candidates and do not claim physical-device validation. The application must retain the portable fallback and must not activate an AOT candidate until its runtime/compiler versions and target device have been validated.

License

The model weights and derivatives are subject to the NVIDIA Open Model License.