Qwen3.8 27B Sparkle 1.0
NVIDIA H200 · vLLM · 65,536 ctx
- generation:
- 165.0 tokens/s
- quant:
- FP8 (FP8)
- kv:
- FP8
mathcodingtool-use
User reports a fine-tune of Qwen3.8-27B, Kvark-Sparkle-1.0, decoding at 165 tok/s versus 171 tok/s for the base model on one H200. Setup is vLLM 0.30 with FP8 weights and FP8 KV cache, MTP speculative decoding with 3 speculative tokens, one user, 64k context cap. Median wall clock is 13.7 s per answer versus 28.1 s for the base, with the gain attributed to 60-67% fewer reasoning tokens. Accuracy is reported as matching the base within the confidence interval on GPQA Diamond, AIME, LiveCodeBench, IFEval and MMLU-Pro.