Latest Content
Cosmos Reason 2 2B icon

Cosmos Reason 2 2B

NVIDIA
Cosmos-Reason2-2B is a multimodal vision-language model optimized for physical AI and robotics. It contains 2.4B parameters, built on a ViT vision encoder and a 28-layer dense transformer with 16 attention heads and 2,048 hidden size. Post-trained from Qwen3-VL-2B-Instruct, it supports up to 256K token context and spatio-temporal reasoning for robotics, video analytics, and data curation. The model enables object detection with 2D/3D localization, common-sense planning, and embodied reasoning, serving as the foundation for AI agents to interpret, plan, and act in complex real-world environments.
TypeVision-Language Model
CapabilitiesText Generation, Instruction Following, Reasoning, Mathematical Reasoning+6 more
Group Release DateJanuary 4, 2026
Links
Blog|HF Model Card
Licensenvidia-open-model-license

Inference Instructions

Deploy and run this model on NVIDIA B200 GPUs using the command below. Copy the command to get started with inference.

CONSOLE
docker run -it --rm 
 --runtime=nvidia 
 --gpus all 
 --ipc=host 
 --shm-size=128g 
 -p 8000:8000 
 -v ~/.cache/huggingface:/root/.cache/huggingface 
 -e HF_TOKEN='YOUR_HF_TOKEN' 
 vllm/vllm-openai:v0.17.1 
 nvidia/Cosmos-Reason2-2B 
  --tensor-parallel-size 8 
 --max-model-len auto 
 --max-num-batched-tokens 65536 
  --gpu-memory-utilization 0.95 
  --max-num-seqs 1024 
  --kv-cache-dtype auto 
  --trust-remote-code
Note

Nvidia Cosmos Reason 2 is a gated model; ensure you have been granted access on Hugging Face and have authenticated your environment using a valid HF_TOKEN.

Model Benchmarks

Each model was tested with a fixed input size and total token volume while increasing concurrency to measure serving performance under load.

ITL vs Concurrency

Time to First Token

Throughput Scaling

Total Tokens/sec vs Avg TTFT

Vultr Cloud GPU

NVIDIA HGX B200

Deploy NVIDIA B200 on Vultr Cloud GPU

Deploy NVIDIA B200 instancesContact Sales