Nanbeige 4.2 3B is producing garbled reasoning output in llama.cpp

LaurentPayot · reddit · 2026-07-28

Nanbeige 4.2 3B is spitting garbled reasoning output in llama.cpp

The user reports that Nanbeige 4.2 3B, running on the latest llama.cpp, produces broken reasoning traces and nonsensical output even for a simple “hi” prompt. Two different agent setups behave badly: one emits repeated </think> fragments and random text, while another seems to loop forever.

The post includes the full launch command with --reasoning on, --reasoning-preserve, a 65,536 context window, and quantized KV cache settings. It is essentially a model/runtime bug report asking whether others are seeing the same issue.

Original post →

More from Models

Models channel →