nvidia

/

Cosmos3-Edge

like 93

Follow

NVIDIA 63.7k

Replace the native cosmos-framework Reasoner example with the release-tested vLLM serving flow.

Changes:

• Add vLLM container and single-GPU serving instructions.
• Replace the offline Reasoner CLI with an OpenAI-compatible image reasoning example.
• Keep the displayed result on default inference; thinking is enabled by default.
• Document the supported OpenAI client extra_body setting for a no-think response.
• Send the local example image as a data URL so the client works with a containerized server.

Compatibility:

• README-only change confined to the Reasoning region.
• Verified to merge cleanly alongside discussion #46, which updates only the Generation region.
• Python example syntax and git diff --check pass.
• The default example was executed successfully against the latest vllm/vllm-openai:cosmos3 image and latest model weights.