Skip to main content

Nemotron Omni

The nemotron_omni family owns the officially documented prepared nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 checkpoint. It implements text, image, acoustic and temporal-video reasoning using native TensorRT Python graph construction and the Model Connect C++ runtime, without ONNX.

Optional Edge-LLM 0.11.0 offload is family-owned. Compatible source or wheel Python tools use the existing CMake native SDK discovery contract. Text/image tasks may use Edge; audio/video tasks select native directly, before Edge preparation. Failed Edge preparation for eligible tasks warns and retries native once. Publication and inference failures do not switch backends.

python -m tensorrt_model_connect build /path/to/Nemotron-Omni \
  --family nemotron_omni --task vision_language_generation \
  --precision fp16 --quantization nvfp4 --max-sequence-length 4096 \
  -o nemotron-omni.bundle

Evidence boundaries​

The native build and all six public task modes execute. The current owning fixture uses independently verified factual grounding/transcription questions; all six pass the unchanged 0.15 edit-distance gate in a replay with the corrected native decoder and retained media plans. This replay is not a fresh full-builder run of every source change; the declared premerge case still builds and executes the complete model. Original free-form image, video and image/audio accuracy failures remain recorded and runnable, not reclassified as captioning passes.

The public Model Connect Edge build and original text/image executions also succeeded: text passed quality, while the free-form image case failed the gate. The revised factual image question has not been rerun through Edge. Direct Edge audio and typed-video execution errors remain separate known limitations. Component tests and scoped factual answers are not catalog-wide qualification.

Small native FP8 and NVFP4 projections execute correctly on Ada; lack of FP4 acceleration alone is not a correctness restriction. TensorRT rejects these FP8-containing graphs on A30. These component observations are not full-model qualification or a family-owned GPU allowlist.

See the family-owned recipe for pinned checkpoint details, public task bindings, capacity/timing contracts, installation behavior and reproducible owning validation.