Cosmos Reason 2 is NVIDIA's open reasoning VLM for physical AI, available in 2B, 8B and 32B sizes post-trained from Qwen3-VL. It reasons over video and images about space, time and physics, can serve as a planning model for embodied agents, and adds 2D/3D point localization and bounding boxes with explanations plus long context up to 256K input tokens. On the Physical AI Bench comparison in the model card, the 8B model scores 56.90 on the Robotics category versus 53.08 for Qwen3-VL-8B-Instruct. Weights are released under the NVIDIA Open Model License (commercially usable); the 8B variant needs at least 32 GB of GPU memory.
| Architecture | Multimodal LLM post-trained from Qwen3-VL (ViT vision encoder + dense Transformer LLM); 8B variant based on Qwen3-VL-8B-Instruct; chain-of-thought reasoning for physical AI. |
|---|
| Params | 2B, 8B (8,767,123,696 parameters) and 32B variants |
|---|
| Inputs outputs | Inputs: text + video (mp4, recommended 4 FPS) or image (jpg), up to 256K input tokens; output: text (reasoning, answers, 2D/3D point localization and bounding boxes). |
|---|
| Training data | Cosmos-Reason1 datasets plus EgoExo4D, PerceptionTest, Language Table, IntPhys, InfLevel, CLEVRER; March 2026 update added VideoPhy2, NVIDIA Physical AI AV and Metropolis data, Hyperism and EgoExOR. |
|---|
| Embodiments supported | Embodiment-agnostic planner/reasoner for robots (humanoids, arms) and autonomous vehicles; used as the reasoning component in GR00T workflows. |
|---|
| License weights | NVIDIA Open Model License (commercial use allowed) |
|---|
| License code | Apache-2.0 (cosmos-reason2 repository) |
|---|
| Hardware requirements | 8B model requires at least 32 GB GPU memory; tested on NVIDIA Hopper/Blackwell (H100, A100 test hardware), BF16, Linux. |
|---|
| Benchmarks claimed | |
|---|
| How to try | huggingface.co/…/Cosmos-Reason2-8B |
|---|
| Sources | |
|---|
Benchmarks, where listed, are as claimed by the vendor.
Illustration
