MiniMax M3 428B — B200 vs B300
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and B300 (NVIDIA Blackwell) on MiniMax M3 428B. 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.
B200 posts 5814 tok/s/chip for $0.08 per million tokens at 121 tok/s/user on MiniMax M3 428B; B300 posts 6001 tok/s/chip for $0.11. B200 is 28% cheaper per token; B300 delivers 3% more tok/s/chip.
Throughput at 235 tok/s/user on MiniMax M3 428B: B200 hits 3194 tok/s/chip, B300 hits 3098. Per-million costs land at $0.15 and $0.20 respectively. B200 is 33% cheaper per token; B200 delivers 3% more tok/s/chip.
B200 / B300 on MiniMax M3 428B at 349 tok/s/user: 1636 / 2067 tok/s/chip, $0.29 / $0.31 per million tokens. B200 is 6% cheaper per token; B300 delivers 26% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:5813.8B300:6001.2 | B200:3193.8B300:3097.7 | B200:1636.0B300:2066.5 |
| Cost ($/M tok) | B200:$0.084B300:$0.107 | B200:$0.151B300:$0.201 | B200:$0.293B300:$0.311 |
| tok/s/MW | B200:3399910B300:3158506 | B200:1867716B300:1630350 | B200:956701B300:1087633 |
| Concurrency | B200:~20B300:~13 | B200:~21B300:~8 | B200:~12B300:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.