Community
DFlash2 speculative decoding now works in llama.cpp for Qwen3.8 Flash Next
A Reddit user reports DFlash2 drafts run about 17% faster on average than MTP on Apple M5 Max, with gains on code and list prompts but a small regression on prose. Draft heads are about 350MB at Q4_0 and vision drafting works.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
PSA: Upgrade Flash Next to DFlash2 if you haven't! +17% on average