Cosmos-Reason1 is a suite of multimodal LLMs, ontologies, and benchmarks designed to imbue AI with physical common sense and embodied reasoning capabilities. Targeting researchers and developers in AI, robotics, and embodied agents, it enables models to generate physically grounded responses through long chain-of-thought reasoning.

How It Works

The models, Cosmos-Reason1-8B and Cosmos-Reason1-56B, undergo a four-stage training process: vision pre-training, general Supervised Fine-Tuning (SFT), Physical AI SFT, and Physical AI reinforcement learning. This approach leverages ontologies for physical common sense and embodied reasoning, coupled with custom benchmarks, to specifically enhance the physical reasoning abilities of multimodal LLMs.

Quick Start & Requirements

• Installation and usage details are not provided in the README.
• Requires access to the models, which are not directly downloadable from the repository.
• Further information is available on the Product Website and through associated Papers.

Highlighted Details

• Two multimodal LLMs released: Cosmos-Reason1-8B and Cosmos-Reason1-56B.
• Focus on physically grounded responses and embodied decision-making.
• Development of ontologies for physical common sense and embodied reasoning.
• Creation of benchmarks to evaluate Physical AI reasoning capabilities.

Maintenance & Community

• Developed by NVIDIA.
• No community links (Discord, Slack, etc.) are provided in the README.

Licensing & Compatibility

• Source code is released under the Apache 2.0 License.
• Models are released under the NVIDIA Open Model License.
• Custom licenses are available upon contact. Commercial use may be subject to the NVIDIA Open Model License terms.

Limitations & Caveats

The README lacks specific instructions for installation, usage, or direct model access, requiring users to consult external resources. Model family details are marked as "Coming Soon."