What causes AI models like Qwen to output Thai in Zork?

Strict prompts suppress English patterns, triggering multilingual fallbacks from training data.

Why can't Gemma or Qwen escape Zork's maze?

They ace single steps but lose state across chains—no true world model or memory.

With tools, yes—but locals expose harness flaws, pushing real agent progress.

🤖 AI Dev Tools

An RTX 5080 whirs. Qwen boots up for Zork. It spits Thai. Local AI agents? Not ready for prime time.

theAIcatchup Apr 08, 2026 3 min read

Local LLMs like Qwen and Gemma flop hard in Zork, scoring near-zero due to state loss and language glitches. 𝕏
Tight prompts trigger Thai/Chinese output—unleash agents carefully. 𝕏
Dynamic state summaries boost play, but small models demand better architectures for true agency. 𝕏

Published by

Ship faster. Build smarter.

Get the best Developer Tools stories of the week in your inbox — no noise, no spam.

Originally reported by dev.to