Cosmos3-Edge reasoning

The cosmos3_edge family implements text, image, video and mixed image/video reasoning for nvidia/Cosmos3-Edge. It is distinct from Cosmos3 Diffusers image/video generation and does not produce trajectories or policy actions.

The family owns optional Edge-LLM 0.11.0 offload and a native TensorRT Python-graph / Model Connect C++ fallback without ONNX. Compatible wheel-installed Python tools and source-installed tools use the same native SDK discovery contract.

Build on the target GPU:

python -m tensorrt_model_connect build /path/to/Cosmos3-Edge \
  --family cosmos3_edge --task image_video_text_to_text \
  --precision fp16 --max-sequence-length 1024 -o cosmos3-edge.bundle

A compatible optional Edge SDK enables FP16 offload. Without one, or if Edge preparation fails, the original request falls back to native construction with a warning for a failed preparation. FP32 selects the pure TensorRT native graph.

Evidence boundaries

Native FP32 matches the independent HF FP32 reference exactly on the tested text, image, video and mixed cases. Native FP16 and Edge FP16 pass the tested chat text/image cases but have recorded video and raw-continuation accuracy failures under the unchanged 0.15 normalized edit-distance gate. These successful build/inference paths remain runnable; they are not claimed as quality-qualified video configurations. Identical-engine direct Edge runs reproduce Model Connect's Edge outputs.

See the family-owned recipe for the exact checkpoint, modalities, capacities, public task APIs, installation contract and validation limitations.

Evidence boundaries​

Evidence boundaries