Qwen3.5 35B hits 10.33 t/s on $300 laptop

Impressive, but the real story is architectural and operational, not just raw model size: Qwen's MoE design, quantization, and speculative decoding are doing a lot of the work here.

• –Q4_K_S plus only about 3B active params makes this much lighter in practice than a dense 35B run
• –ik_llama.cpp and MTP speculative decoding are part of the performance story, so backend choice clearly matters
• –The laptop is tuned hard with core pinning, performance mode, and thermal constraints, which makes reproducibility across consumer hardware shaky
• –Hitting 90C suggests this is near the edge of what the machine can sustain, so sustained throughput may be less stable than the headline number implies
• –For local LLM users, this is a useful reminder that memory bandwidth, thermals, and inference stack quality can matter as much as CPU specs