Speculative decoding draft models for Qwen3-8B compared: WhiFlash vs Tengyunw benchmarks
maxsloef · x · 2026-11-12
A thread surfaces multiple speculative-decoding draft models for Qwen3-8B. WhiFlash/Qwen3-8B-EAGLE3, trained by Samsung AI Center Cambridge on EAGLE-3 with a fixed 32K draft vocabulary, runs losslessly with SGLang (research-only license). The shared benchmark table compares WhiFlash, Tengyunw, RedHatAI and AngelSlim: WhiFlash leads with thinking off (0.750 top-1 / 2.47 mean correct run), Tengyunw leads with thinking on (0.711 / 1.96).
Related event: Benchmark Compares Draft Models for Qwen3-8B Speculative Decoding(2 posts)→
More from Infra
- NVIDIA plunges 9% after hours on report board may fire Jensen Huang — djcows · 2026-11-12
- Vmax teases Campaign: self-play RL task generation taught agents to use Modal SDK — soleio · 2026-11-12
- JITEngine synthesizes custom LLM inference engines for Apple Silicon, up to 2.7x faster — CShorten30 · 2026-11-12
- LeapQuant: training-free quantization makes linear attention up to 4.25x faster — tianyin_xu · 2026-11-12
- Microsoft's MAI Transcribe 2 sees 12x usage growth, hitting 8.5M weekly transcriptions — mustafasuleyman · 2026-11-12
- gVisor joins CNCF with Modal as co-maintainer after upstreaming GPU and training support — josh_wills · 2026-11-12