Benchmarks

Every run's scripts and results are in the benchmarks repo on GitHub. All of them ran on this machine unless a post says otherwise:

  • GPU: RTX PRO 6000 Blackwell Workstation Edition, 96 GB, 400 W power limit
  • CPU: AMD Ryzen 9 7950X
  • RAM: 250 GiB DDR5
  • OS: Ubuntu 24.04, NVIDIA driver 580 (CUDA 13.0)

Full list: hardware.md .

Date Benchmark Data
76–82% of DFlash2 rejects on Qwen3.8-27B NVFP4 happen where the target's top token is under 90% likely repo
Inkling 2 Mini reaches 2.2x Qwen3.8-27B's output rate on a 120k prompt in 41.5 GiB, 1.2x at chat, and scores lower on GSM8K repo
Stock Qwen3.8-27B NVFP4 scores 66.09 on the Decision Index public vision suite; 29 of 62 board entrants score higher repo
Qwen3.8-27B NVFP4 decisions on images: a temperature fitted on image sets carries to other image sets of the same question type repo
A decision temperature pooled from other text datasets keeps 49–97% of the in-domain gain on Qwen3.8-27B NVFP4 repo
Per-dataset client-side temperatures improve Qwen3.8-27B and Swift 1.5 NVFP4 decision probabilities repo
Swift 1.5 NVFP4 vs stock Qwen3.8-27B: no clear quality gap, 14–19% fewer thinking tokens at medium repo
Qwen3.8-27B with DFlash2 matches or beats Flash-Next on prompts up to 40k tokens, at half the GPU memory repo
DFlash2 8 led my Qwen3.8-27B drafter tests on one RTX PRO 6000 repo
RadixArk's NVFP4 Qwen3.8-27B served 32–37% more tokens/s than Qwen's FP8 on SGLang repo