Benchmark of unsloth/Qwen3.6-27B-NVFP4 on 4x RTX 5060 Ti

A user benchmarks the new unsloth/Qwen3.6-27B-NVFP4 model using vLLM on a setup of four RTX 5060 Ti GPUs to evaluate prefill performance and Time to First Token (TTFT) across varying concurrency levels.

• The test environment uses pipeline parallelism of 4, 64GB VRAM per card, and max batch tokens of 16K.
• At 1 concurrency, mean TTFT is 3291.57 ms with a total token throughput of 211.95 tok/s.
• Increasing to 4 concurrency raises mean TTFT to 3864.81 ms and total token throughput to 647.57 tok/s.
• At 8 concurrency, mean TTFT increases to 5279.76 ms while total token throughput reaches 1023.60 tok/s.
• With 12 concurrency, mean TTFT jumps to 14441.53 ms and total token throughput is 1047.73 tok/s.
• At 16 concurrency, mean TTFT rises sharply to 37817.92 ms with a total token throughput of 1061.13 tok/s.

The benchmark illustrates how prefill latency and end-to-end latency scale significantly as concurrency increases on this specific hardware configuration.