ThinkingCap squeezes 38.5% fewer thinking tokens out of Qwen3.8-Flash-Next at ~1pp accuracy cost
Thom_Wolf · x · 2026-10-23
BottleCap AI released the fourth model in its ThinkingCap series, a token-efficient version of Qwen3.8-Flash-Next (360GB bf16, served on four H200s).
- Thinking tokens cut 38.5% on average across six benchmarks, up to 51% (RealWorldQA); the enterprise PRO edition cuts 60% on average and up to 75.8%
- Average accuracy drops only 1.03pp (89.8 → 88.8), with the gap within noise on 4 of 6 benchmarks
- Same sampling settings (thinking on, temperature 1.0, topp 0.95) and vLLM serving stack — a drop-in replacement
- Weights and GGUF are on Hugging Face, with a blog post detailing the training objective
More from Models
- SubQ Snap 1.1 ties for first on Artificial Analysis Decision Index with 384K context — ArtificialAnlys · 2026-10-23
- PolyAI unveils Dialog-RSN-1: audio-native voice model answering in under 300ms — matthen2 · 2026-10-23
- Konrad Kording: little evidence that chatting with AI personalizes it — KordingLab · 2026-10-23
- Meta paper ditches tokenizers: distilled 1B byte models outperform token-based peers — mtizard · 2026-10-23
- Terminal-RSI lifts Qwen3.8-27B from 51.7% to 86.52% on Terminal-Bench, plus 70 tok/s on M5 Max via DFlash 2 — udmrzn · 2026-10-23
- SubQ Snap 1.1 Launches: 384k-Context Decision Model Tied for #1 on AA Decision Index — ryanallred · 2026-10-23