Reproducible serving recipes and customer-reference benchmarks, one folder per accelerator. Every number is labelled measured (our receipt), community-reported, or inferred.
| model | GPU | engine / image | c=1 | c=16 | c=128 | $/M output tokens | receipt |
|---|---|---|---|---|---|---|---|
| Qwen3.8-Flash-Next NVFP4 (nvidia) | 2x B200 TP2 | vLLM nightly 2026-09-06 | pending | 1,501 | 5,256 | 0.66 | b200/qwen3.8-flash-next |
| GLM-5.3-Flash NVFP4 (LibertAIDAI @11d73216) | 2x B200 TP2 | vLLM nightly 2026-09-06, Marlin MoE | pending | pending | ~2,300 (thinking on, inferred) | ~1.5 | b200/glm-5.3-flash |
| DeepSeek-V4.1-Flash (bf16/fp8, 475.3 GiB) | 4x B200 TP4+EP4 | SGLang @c4ca6511, DSPARK spec | 331.0 | 1,724.4 | not reachable | 4.03 (at c=16) | b200/deepseek-v4.1-flash |
Pricing basis: Modal list price $6.25/B200-hour. See docs/methodology.md.
Notes on the DeepSeek-V4.1-Flash row (all measured): the c=16 figure comes
from the knee sweep at admission cap 256, not from the two boots at cap 32, which
disagree by 1.46x at that level — see the recipe. c=16 is this model's throughput
peak, not a mid-sweep point — aggregate falls to 667 tok/s at c=32,
so $4.01/M is a floor rather than a trend. c=128 is not reachable on this node:
--max-running-requests 512 will not boot because the sliding-window pool wants
31.56 GB against 18.88 GB free. A 262,144-token prompt kills the server on a model
advertising a 1M context, while 120,177 tokens serve at 51,527 tok/s.
b200/ h100/ a100/ a10g/ b300/ vera-rubin/
<model>/ notebook.ipynb recipe.md receipts/
docs/methodology.md
Related: club-170hx, club-dgx-spark (owned-hardware communities).