JANGQ-AI

/

GLM-5.3-Flash-EXL3

like 0

Follow

JANG_Q - GGUF for MLX 385

fix: default clear_thinking=true in chat_template.jinja + add repetition_penalty=1.1 default in generation_config.json — mitigates known GLM-5.3 reasoning-loop degeneration on multi-turn agent workloads (SGLang #36669, opencode #45533). Upstream Z.ai default `clear_thinking=false` replays prior reasoning into subsequent turns, priming a decode attractor. Z.ai's own docs recommend `clear_thinking=true` for chat scenarios — bakes it in.