VibeusDFlash2 for Qwen3.8-Flash-Next now works in llama.cpp v0.8.0. In... - Vibeus
DFlash2 for Qwen3.8-Flash-Next now works in llama.cpp v0.8.0. In returnity’s M5 Max test, the geometric mean rose from 75.1 to 87.9 tokens/s versus MTP: +17%.
With the Q4_K_XL target, code went from 80.5→95.1 tokens/s (+18%), list output 91.4→127 (+39%), prose 53.8→51.9 (−3%), and the roughly 6K-context task improved 19%. Acceptance was MTP→DFlash2: code 94→84%, lists 98→98%, prose 78→64%, context 97→90%.
Both drafters were tuned after a parameter sweep on Apple Silicon: MTP n-max 3, DFlash2 n-max 5, both p-min 0.75. Different hardware may need different settings. At 25K context, DFlash2 roughly tied MTP, while MTP’s speed stayed nearly flat.
Q4_0 and Q8_0 draft heads showed no measurable speed or acceptance difference in these tests. Q4_0 is about 350 MB, versus 2.2 GB for the MTP head used — but that does not become a VRAM saving because DFlash2’s compute buffer offsets it.
Image prompts also worked with coherent responses and good acceptance, unlike DFlash2 for Qwen3.8 27B in this comparison. Prose remains the exception to the speedup; these are the tester’s results, not universal gains.
