🚨 you don't need a full thinking budget to keep Qwen-level answers BottleCap j… · @QwenAIFans · TwKit

@QwenAIFans

🚨 you don't need a full thinking budget to keep Qwen-level answers BottleCap just shipped ThinkingCap for Qwen3.8-Flash-Next — drop-in, same sampling, ~40% less thinking. Hard numbers (6 benches, same harness): • thinking tokens −38.5% avg (up to −51% on RealWorldQA) • accuracy 89.8% → 88.8% (−1.03pp) • AIME 2026: 16.7k → 9.3k thinking tokens (−44%) Caveat: bf16 still wants ~4×H200 (~360 GB). NVFP4 cuts that to ~132 GB. via BottleCap AI 👇 https://bottlecapai.com/post/thinkingcap-qwen3-8-flash-next/

曝光 160 · 评论 0 · 点赞 0 · 书签 1