Kimi K2.5/K2.6 1T — B300 vs GB200 NVL72
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and GB200 NVL72 (NVIDIA Blackwell) on Kimi K2.5/K2.6 1T. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
B300 / GB200 NVL72 on Kimi K2.5/K2.6 1T at 70 tok/s/user: 723 / 1324 tok/s/GPU, $0.89 / $0.46 per million tokens. GB200 NVL72 is 92% cheaper per token; GB200 NVL72 delivers 83% more tok/s/GPU.
Around the middle of the 45–146 tok/s/user interactivity band, at 96 tok/s/user on Kimi K2.5/K2.6 1T: B300 runs 465 tok/s/GPU at $1.37/M tokens, GB200 NVL72 runs 665 at $0.97/M. GB200 NVL72 is 42% cheaper per token; GB200 NVL72 delivers 43% more tok/s/GPU.
Setting 121 tok/s/user as the target on Kimi K2.5/K2.6 1T, B300 produces 286 tok/s/GPU ($2.25 per million tokens) and GB200 NVL72 produces 181 ($3.39). B300 is 51% cheaper per token; B300 delivers 58% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | B300:723.3GB200 NVL72:1323.8 | B300:465.4GB200 NVL72:664.7 | B300:286.1GB200 NVL72:181.0 |
| Cost ($/M tok) | B300:$0.893GB200 NVL72:$0.464 | B300:$1.371GB200 NVL72:$0.968 | B300:$2.247GB200 NVL72:$3.394 |
| tok/s/MW | B300:333315GB200 NVL72:630368 | B300:214492GB200 NVL72:316546 | B300:131822GB200 NVL72:86184 |
| Concurrency | B300:~21GB200 NVL72:~372 | B300:~10GB200 NVL72:~136 | B300:~4GB200 NVL72:~16 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.