Qwen3.5 122B A10B at UD-Q2_K_XL fits in 64GB RAM and improves internal knowledge

A user demonstrates that Qwen3.5 122B A10B quantized with UD-Q2_K_XL can fit into 64GB of system RAM, replacing the previous best option for that constraint.

• The model achieves higher quality responses and better internal knowledge compared to Qwen3 Next 80B at UD-Q4_K_XL.
• Generation speed drops to approximately 2.9 tok/s on DDR4 CPU, though DDR5-5600 could raise this to 6 tok/s.
• The architecture uses 10B active parameters via MTP, whereas Qwen3 Next 80B uses only ~3.5B active parameters.
• REAP variants of the model were found to be inferior and did not resolve slow prompt processing speeds.

This configuration allows users with limited memory to access a larger model's capabilities, though the slow prompt processing makes it unsuitable for agentic workloads requiring quick information lookup.