Tag
A user tweets about downloading the second K3 model's full FP16 weights and storing them, jokingly asking others to leave their doors unlocked.
The author shares surprising results from systematic tests on how different quantization levels (e.g., Q4_K_M, Q5_K_M) affect model capabilities separately, showing that math accuracy degrades more than knowledge tasks, and calls for more rigorous testing on context decay across quant levels.
The author verified that converting the Qwen3.6 27B model weights from BF16 to FP16 does not cause numerical overflow, and pointed out that FP16 has higher mantissa precision, explaining why quantized versions use FP16 instead of keeping BF16.