Cosmos 3 and Qwen 3.5 benchmarked on Jetson T5000.

YUAN benchmarked NVIDIA Cosmos3-Nano, Cosmos3-Edge, and Qwen3.5-9B on the NVIDIA Jetson T5000 platform. The tests compare vision-language reasoning accuracy, F1 score, and token throughput under single-stream and multi-stream workloads.

Accuracy against throughput

At 49 concurrent streams, Cosmos3-Nano sustained 525.01 total tokens per second. Qwen3.5-9B sustained 215.37 tokens per second. In the same vLLM and NVFP4 configuration, measured accuracy was 96.31 percent for Cosmos3-Nano and 97.44 percent for Qwen3.5-9B.

Qwen3.5-9B recorded the highest measured accuracy and F1 score at 97.44 percent and 91.15. Cosmos3-Nano recorded 96.31 percent accuracy and an F1 score of 88.14 while processing more total tokens per second in the 49-stream test. The deciding trade-off is model quality versus aggregate throughput for the tested workload.

Full-precision comparison

Cosmos3-Edge does not yet support NVFP4 quantization or the vLLM runtime. YUAN therefore ran all three models in FP32 on the Hugging Face Transformers runtime. In that test, F1 scores were 90.32 for Qwen3.5-9B, 88.32 for Cosmos3-Nano, and 87.40 for Cosmos3-Edge.

Single-stream throughput in the FP32 test was 7.37 tokens per second for Cosmos3-Edge, 3.83 for Cosmos3-Nano, and 2.95 for Qwen3.5-9B. These results give developers a separate baseline for single-camera or power-constrained systems.

Multi-camera deployment path

YUAN combines the Jetson T5000 with its video capture hardware, SmartNVR, and SmartVMS software for on-site multi-camera analysis. Teams should validate the same model, runtime, precision, stream count, and input resolution under their own workload before selecting a deployment configuration.