Qwen3.5-35B Matches Q4 on MI50s

This reads less like a surprise model win and more like a reminder that local inference performance is often limited by kernels, memory movement, and device topology, not just bit width.

• –Q8_0 keeping pace with Q4_K_XL on generation suggests the bottleneck is not purely arithmetic throughput.
• –The prefill jump on two GPUs shows where parallelism still matters: prompt processing benefits far more than token-by-token decoding.
• –MI50-era AMD cards are exactly where inference stacks tend to be least polished, so quantization gains can get swallowed by software inefficiency.
• –For local model runners, this is a useful reminder to benchmark `prefill` and `decode` separately before choosing a quant level.
• –Qwen3.5-35B-A3B still looks attractive for multi-GPU local deployment, especially if you can afford a higher-quality quant without losing real-world speed.