DFlash2 drafter lifts Qwen3.8-Flash-Next speed by 17%

A user's tests show enabling the DFlash2 drafter for Qwen3.8-Flash-Next in llama.cpp v0.8.0 raises decode speed by about 17% geometric mean: +18% on code, +39% on lists, -3% on prose. The DFlash2 draft head is only ~350MB versus 2.2GB for MTP and supports vision input.

• Eagle-4 Released: Faster than DFlash2 & DSpark2026-10-21
• User builds 256GB VRAM rig with 8 Radeon Pro V620s2026-10-08
• User Shares Strata Speedup for Local LLMs but 64GB RAM Constraint2026-10-04
• Qwengram-0.8B: n-gram memory transfer cuts perplexity 5.05%2026-09-25
• Qwen3.8-Flash-Next Hits 65 Tokens/Second on 12GB VRAM2026-09-24