Upgrade Flash Next to DFlash2 in llama.cpp for ~17% Speed Boost
With llama.cpp now supporting DFlash2 for Qwen3.8-Flash-Next, local decode speeds improve by an average of 17% while draft head size shrinks to ~350MB.
According to community tests, llama.cpp now supports DFlash2 for Qwen3.8-Flash-Next. Benchmarks on Apple Silicon (M5 Max) show:
- Performance: Geometric mean speed increased by ~17% compared to MTP. List prompts saw a 39% boost (from 91.4 tps to 127 tps), while prose experienced a minor -3% regression.
- Optimal Parameters: Found to be
n-max = 5,p-min = 0.75for DFlash2. - Size Reduction: The Q4_0 DFlash2 draft head is only ~350MB, significantly smaller than the 2.2GB GGML MTP head.
- Vision Support: Vision-enabled prompts work properly with coherent responses and good acceptance rates.
Source: https://www.reddit.com/r/LocalLLaMA/comments/1xjlgm4/psa_upgrade_flash_next_to_dflash2_if_you_havent