Porting a Qwen 3.6 ThinkingCap LoRA to 3.8 cuts thinking tokens from 33k to 20k
CapsAdmin · reddit · 2026-08-17
A user loaded ThinkingCap LoRAs trained for Qwen 3.6 directly onto Qwen 3.8. Since both share the same architecture, the LoRA applies technically — and in limited testing it "somewhat works."
Using the classic "SVG of a pelican riding a bicycle" prompt with fixed seed, four settings were compared: xhigh, xhigh + rank64 LoRA scaled at 30, xhigh + the original LoRA at scale 1, and medium. The ThinkingCap LoRA reduces thinking tokens from 33k to 20k with comparable (slightly less detailed) output; raising LoRA strength cuts tokens further but degrades quality. medium uses only 3500 tokens but with a noticeable quality drop, leaving ThinkingCap somewhere in between.
The author wrote a Python script that runs llama-cli and exports results to self-contained HTML files embedding the logs, chat turns, launch parameters, and the script itself for full reproducibility. He's unsure whether it genuinely works or just degrades the model into doing less, and invites verification — speculating that merging the weights might work even better.
More from Models
- DFM Mimir v1: Open 1B Model Achieves SOTA Danish Performance — SDU-Denmark · 2026-08-17
- Ling-3.0-flash Runtime Path: Running on One DGX Spark — Kanu-animallover · 2026-08-17
- System prompts beat user prompts: taming verbose Claude Opus 5 — IndyDevDan · 2026-08-17
- Intern-S2-Mobius: Decoupled Knowledge and Reasoning Model — pmttyji · 2026-08-17
- Mimir: 1.7B model claims to beat Qwen and Gemma — ZookeepergameCool173 · 2026-08-17
- Tencent's EVIE Model Tops ViDoRe Benchmarks, Cuts Vector Storage Costs by 32x — jacek2023 · 2026-08-17