Benchmarks & ComparisonsLatest6 min read
Ripple-65B on One H200: 43 Tokens per Second for One User, 1,197 at 64
Ricursive's Ripple-65B Preview, the open coding model that reports scores above 500B to 975B mixture-of-experts models, runs on a single H200. We served it with vLLM in FP8 on a rented H200 NVL and measured it from 1 to 128 concurrent requests, at 16k and 60k of context, and what it costs per million tokens. Plus the two things that stop it from starting today.
- 43 tok/s
- one request, a 65B dense model on one H200
- 1,197 tok/s
- in total at 64 concurrent requests
- 62.6 GiB
- of weights in FP8, with 486k tokens of KV cache left
- $0.81
- per million output tokens at 64, at $3.50/h


























































































