Cosmos Reason

Cosmos Reason is an open, customizable reasoning vision-language model (VLM) developed by NVIDIA as part of the Cosmos platform for Physical AI and robotics. It enables robots and vision AI agents to reason using prior knowledge, physics understanding and common sense in order to understand and act in the real world. The model understands space, time and fundamental physics, and can serve as a planning model that reasons about what steps an embodied agent might take next.

Architecture

A multi-modal LLM consisting of a Vision Transformer (ViT) as the vision encoder and a dense Transformer as the language model. Network architecture: Qwen2.5-VL-7B-Instruct — Cosmos Reason1-7B is post-trained on top of Qwen2.5-VL-7B-Instruct. Parameter count: ViT 675.76M + LLM 7.07B + projection layer 545M (about 8B total). Video/image is converted into tokens by the vision encoder and a projector, combined with the text prompt and fed into the core model, which uses chain-of-thought to answer step by step.

Training

The model is post-trained on physical common-sense and embodied-reasoning data using supervised fine-tuning (SFT) and reinforcement learning (RL). It uses chain-of-thought reasoning to understand world dynamics without human annotations. Training data includes RoboVQA, BridgeDataV2, AgiBot, HoloAssist and autonomous-vehicle (AV) driving data collected and annotated by NVIDIA.

Applications

Data curation and annotation (automating high-quality annotation of massive datasets), robot planning and reasoning (the brain for vision-language-action models — a robot breaks complex commands into tasks and executes them), and video analytics AI agents (extracting insights and root-cause analysis over large volumes of video). The model is ready for commercial use.

Availability

Cosmos Reason1-7B weights are publicly available on Hugging Face under the NVIDIA Open Model License (commercial use permitted). Post-training code is in the nvidia-cosmos/cosmos-reason1 repository. Runtime: vLLM. Tested hardware: H100, A100, GB200 (NVIDIA Hopper / Blackwell), BF16 precision. The next generation — Cosmos Reason 2 — was released in October 2025. The Cosmos family was subsequently merged into the Cosmos 3 omni-model (COMPUTEX 2026).