Hobbyist's 'Weight Space Programming' on Qwen 3.8 cuts token usage up to 58% with no accuracy loss
Xananique · reddit · 2026-11-07
A Reddit user demonstrates what they call "Weight Space Programming": starting from the HauHauCS Qwen 3.8 Balanced Abliteration checkpoint, they used ChatGPT to iteratively rewrite model weights over 6 hours, producing a 4-bit variant dubbed Calmer-Qwen-3.8, published on Hugging Face (Calmer-V2-Qwen3.8-27B-4bit) and GitHub.
Screenshot data shows the merged layers-18-49 variant cut tokens on a five-question all-correct panel from 23,081 to 18,461 (-22.4%) and on topology-wrong answers from 39,407 to 18,648 (-58.1%), with no loss in correctness. Repeated aphorism-style sentences dropped from 38 to 1. The author describes the model as "less anxious and straight to the point," though one long-reasoning puzzle remained unfinished and accuracy gains aren't yet established. Experimental data analysis is still to come.
More from Models
- Grok bots flood social feeds with UGC ads as xAI pushes B2C, Reddit asks if they're any good — digitalbananax · 2026-11-07
- Inkling 2 Small pretrained on 8.4T tokens in 68 hours, natively in NVFP4 — teortaxesTex · 2026-11-07
- TTS leaderboard shakeup: accent alone moves Elo, HeyGen Voice tops overall board — ArtificialAnlys · 2026-11-07
- InstantHMR open-sourced: real-time 3D human mesh recovery in 5.5ms per person on phone NPU — PurzBeats · 2026-11-07
- Bittensor week: subnets hit model Pareto frontier, GPU rentals hit $2.32M in October — markjeffrey · 2026-11-07
- Local ripple-65B scores 38/49/27 on agentic coding: errors encountered, not draft quality, decided the grade — Ok-Brush8377 · 2026-11-07