vLLM 0.32.0
NVIDIA B300 x4 · tp4 × pp1- Speed per user (c=1)
- 192.7tok/s
- Time to first token
- 33ms
- Peak throughput
- 12,730tok/s @ 256
- Per GPU at peak
- 3,183tok/s
- Cold start
- 1,700s
- Context tested
- 8,192tokens
Measured 9 Nov 2026 · Throughput sweep
Measured 9 Nov 2026 · Throughput sweep
ModelPerf Pro