Cosmos3 Reasoner NIM fails with 400 Bad Request fetching cosmos3-nano-reasoner profile file maps

AI & Data Science NGC Catalog Feedback
,
1

Title:
Cosmos3 Reasoner NIM fails with 400 Bad Request while fetching cosmos3-nano-reasoner profile file maps on RTX PRO 6000 Blackwell

Body:
I am testing NVIDIA Cosmos3 Reasoner NIM locally for an Isaac Sim / robotics visual QA workflow.

Hardware / environment:

  • GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
  • VRAM: 96 GB
  • Driver: 580.88
  • Host OS: Windows 11
  • Docker Desktop with NVIDIA GPU passthrough
  • NIM image: nvcr.io/nim/nvidia/cosmos3-reasoner:latest
  • Model size: nano
  • Served model selected by NIM: nvidia/cosmos3-nano-reasoner
  • Cache: clean cache path, no stale containers

Launch shape:
docker run --rm --name nvidia-cosmos3-reasoner
–gpus all
–ipc host
–shm-size=32GB
-e NGC_API_KEY
-e NIM_MODEL_SIZE=nano
-v C:\nim-cache\cosmos3-reasoner-clean:/opt/nim/.cache
-p 8000:8000

nvcr.io/nim/nvidia/cosmos3-reasoner:latest

Observed behavior:
The NIM starts successfully, detects Nano 8B, and begins profile selection. It then fails while fetching model profile file maps from NGC. The same failure appears for nvfp4, fp8, and bf16 Nano profile variants.

Example errors:
HTTP status client error (400 Bad Request) for:

https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/cosmos3-nano-reasoner/modelopt-nvfp4-full-quantize-final_format_fix/files

HTTP status client error (400 Bad Request) for:

https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/cosmos3-nano-reasoner/modelopt-fp8-final_format_fix/files

HTTP status client error (400 Bad Request) for:

https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/cosmos3-nano-reasoner/bf16-final/files

What I already tried:

  • Removed old stopped Cosmos3 containers.
  • Used a fresh cache directory.
  • Used nvcr.io/nim/nvidia/cosmos3-reasoner:latest.
  • Previously also tried 1.7.0 and 1.7.
  • Used NIM_MODEL_SIZE=nano.
  • Used --ipc host and --shm-size=32GB.
  • Confirmed the container starts and reaches NIM profile selection.

Question:
Is Cosmos3 Nano Reasoner NIM currently supported for local deployment on RTX PRO 6000 Blackwell Workstation Edition?

If yes:

  • Is there an additional NGC entitlement, license acceptance, org/team access, or model access step required for cosmos3-nano-reasoner?
  • Is there a recommended environment variable to force a supported profile for RTX PRO 6000 Blackwell?
  • Are the listed fp8/nvfp4/bf16 profile file-map URLs expected to be accessible for public NIM users?

If no:

  • Is hosted NVIDIA Build/API currently the recommended path for Cosmos3 Nano Reasoner?
  • Is there an ETA or supported hardware/profile matrix for local Cosmos3 Reasoner NIM on workstation GPUs?

Thank you.

2

I solved this- My Personal API key was wrong type. now it works ok!