Llama 3.2 1B on QCS6490: 1,024 tok/s prefill, 16.4 tok/s decode

The two runs share more than a chip. The recipe was identical: read the pipeline's own compile parameters out of a failed export job, drop the single problematic flag, and pin the input shapes the script forgets to fix. That last step matters — the author notes the first Qwen3 attempt died because shapes were left dynamic. With six compiles, one link and six profiles on an RB3 Gen 2 all passing, the bug remains inside the script while the workaround routes around it.

What the author is really building is a series, not a single result. The comparison table now holds two rows at w4a16 and context 512: Qwen3-0.6B at 1,555 tok/s prefill and 24.6 tok/s decode, and Llama 3.2 1B Instruct at 1,024 and 16.4. The stated method is what makes the rows comparable, and the author says he would rather publish numbers bundled with their method than larger numbers without one. The next entries in the queue — Qwen3-1.7B and Llama 3.2 3B — would turn a two-point comparison into a trend, moving from extrapolation toward measured answers on what this chip runs and how fast.