A new draft head called DFlash2 has been released for the Qwen3.8-Flash-Next model, offering significant speed improvements over the previous MTP draft head. Benchmarks show an average increase of 17% in tokens per second, with some categories seeing up to a 39% boost. While DFlash2 is notably smaller in file size, its performance advantage is most pronounced in non-prose tasks, and it also supports image generation, unlike some previous versions. AI
IMPACT Offers significant decode performance gains for specific LLM tasks, potentially speeding up inference for users.
RANK_REASON Release of a new draft head for an existing model, with performance benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
