Original writeup: 37 AI novel runs show smaller context windows win
lmoroney · x · 2026-10-12
This is the original writeup behind the earlier tweet on Jessica K. Murray's 37-novel local-model experiment. It details how 8k context beats 32k (repetition 42.8%→0.5%), a 6 GB card running gemma4:26b writes a clean 65k-word book, only 1 of 5 quality tools ranks correctly, chapter-length decay signals looping, and the recommended setup: gemma4:26b with context explicitly set to 8k on both 6 GB and 12 GB machines.
Related event: Test of 37 AI Novels Finds 8k Context Cuts Repetition to 0.5%(2 posts)→
More from Models
- Step 5 Preview tops OpenRouter trending; 3D island demo rivals Opus at lower cost — GCWebDesigner · 2026-10-12
- OpenRouter token traffic explodes from 2T to 379T/month, open-weight models at 75% — Beth_Kindig · 2026-10-12
- Gadsby-4B Hits Hugging Face: Constrained Decoding Turns Gemma 3 4B Into a Lipogram Model — cephaloform · 2026-10-12
- 50 AI Models Have Played Cowcraft, an Online Limit-Testing Arena for LLMs — djcows · 2026-10-12
- After days of testing, dev ditches sol 6.1 and astra 6 for agentic work, returns to Opus — NERDDISCO · 2026-10-12
- "Flash" isn't fast anymore: new Flash-tier models deliver only ~50tps — TheAIPU-guy · 2026-10-12