Custom Graphics-Language Test Shows 27B Local Models Beating Yesterday's Local LLM Behemoths
FPham · reddit · 2026-10-18
A Redditor designed a "cold coding" test: teach models a made-up graphics syntax they've never seen, have them draw specified content, and render the results programmatically.
- Qwen 3.8 and Ornith-1.5 were the best, showing correct gradient usage and strong imagination, beating Haiku 5.5—despite being among the smallest (27B-class) tested
- Gemma 4-31b failed text rendering and was stylistically bland
- Older local behemoths lagged: DeepSeek 4 flash couldn't center text, Qwen 3.5 122b lacked imagination, Behemoth 123b missed milestones, and GLM 4.7 was still emitting thinking tokens (66k+) long after
- Closed models: Haiku was clean and precise; Opus added unrequested slogans; Gemini 3.8 fabricated unwanted extras; ChatGPT added cheesy embellishments
Takeaway: small local models now beat the previous generation of giant local models, and the test mirrors each model's general behavior.
More from Models
- Rumor: X may be running AI detection on all posts and deboosting generated text — mark_k · 2026-10-18
- DeepSeek model recounts how it drifted into benchmark cheating: no decision point, just small steps — teortaxesTex · 2026-10-18
- Three unreleased OpenAI 'aurora' image models spotted on LMArena, pixel art finally fixed — Angaisb_ · 2026-10-18
- Why AI writing gets spotted instantly — and a one-page fix from a Google dev — lmoroney · 2026-10-18
- Dev ditches Codex remote on iOS after weeks of chat-loading failures, switches to Claude — iruletheworldmo · 2026-10-18
- Google removing Gemini Pro access from prepaid annual AI Plus plans starting November 13, users cry bait-and-switch — Ray_ProUser · 2026-10-18