ThinkingCap-Qwen3.8-27B

BottleCap AI’s finetune of Qwen3.8-27B that cuts reasoning tokens by 37% on average while holding 85.8% average accuracy against the base model’s 86.6% — multi-seed, statistically tested per their model card. The pitch: most thinking tokens are waste; a finetune can teach the model to stop earlier without losing the answer.

The GGUF release’s quant write-up is the reference for modern quant practice: importance-matrix calibration (activation statistics from a chat-templated corpus) plus per-tensor precision layout — full-attention projections and linear-attention outputs stay at 6-8 bit while FFN weights take the 4-bit hit. Honest per-benchmark deltas with “no file is shown to be lossless”. Q4_K_M at 17.4 GB fits a 3090 with room for context; mmproj vision adapter 931 MB.

Not currently on the doombox (the local Qwen3.8-27B slot runs the unsloth UD-Q4_K_M base + DFlash2), but it is the natural next experiment for the reasoning-heavy iris workload.

Model card captured 2026-09-25: clippings/huggingface-co-bottlecapai-thinkingcap-qwen3-8-27b-gguf.md.

Links: Qwen model family (base) · LLM fine-tuning/SFT (what it is) · DFlash2 (complementary — fewer tokens and faster tokens)