Qwen3.6 35B (3B active)

User reports Qwen3.6 35B A3B coder at 24.5 tok/s generation and ~118 tok/s prompt processing on a £299.99 Lenovo IdeaCentre Mini with Snapdragon X1-26-100, 32GB RAM, CPU only. Setup is llama.cpp with Q4_0 quant, 512 token prompt, 8 threads. The GPU is an Adreno X1-45 and the NPU is unused, so this is all CPU. With MTP the coder reaches 29-33 tok/s. A custom ARM port of Janas (MoE expert streaming from NVMe) gives 17.2 tok/s plain and 29-31 tok/s with MTP for the coder, and runs an 80B model at 9.8 tok/s plain and 17-18 tok/s with MTP. Gemma 26B runs at 29.5 tok/s in llama.cpp and 23.5 tok/s in Janas. The user also squeezed the coder's expert cache to 12GB and still got 24 tok/s.