64GB-RAM user finds Qwen3.5 122B beats Qwen3 Next 80B despite lower precision
A user running on 64GB of regular RAM compares two setups. Previously they ran Qwen3 Next 80B at UD-Q4_K_XL , getting 8.5 on DDR4 with solid and decent internal knowledge.
Switching to Qwen3.5 122B (10B ) compressed down to UD-Q2_K_XL -- a much lower precision level -- still fits within the same 64GB and, surprisingly, produces noticeably better response quality and internal knowledge than the higher-precision Qwen3 Next 80B. The tradeoff is speed: generation drops to about 2.9 , and slows heavily since both models run on CPU.
With faster DDR5-5600 memory, the user expects generation speed to reach roughly 6 , which they consider usable. They attribute the quality gain to Qwen3.5 122B's standard packing knowledge more effectively than the comparatively outdated sparse MoE design used in Qwen3 Next 80B.
Key points
- Qwen3.5 122B A10B at UD-Q2_K_XL on 64GB RAM gave better response quality than Qwen3 Next 80B at UD-Q4_K_XL
- Speed dropped to 2.9 tokens/sec on DDR4, versus 8.5 tokens/sec for the previous setup
- Faster DDR5-5600 RAM is expected to push generation speed to around 6 tokens/sec
- The quality gain is attributed to a standard packing knowledge more efficiently than the older sparse MoE design