Decode · tok/s
195.8
Prefill
3944
tok/100W
34.1
Qwen3.8-Flash-Next 125B running on one RTX 5090 + 64 GB RAM (IQ3_S, Strata): 22,047 tokens in 112 s 195.8 tok/s generation 3,944 tok/s prompt ~75% draft acceptance (MTP) One prompt, and it one-shot a playable Three.js neon lane runner with bloom, coins, particles and touch controls.
First-person 5090 + 64GB. Rank is the quoted 195.8 generation for the long one-shot. Prefill is the quoted 3,944. Separate from the 142 (128GB) and 104 (256K) IQ3_S pairs on the same chip.
