Qwen3.5 on M4 Pro: dated precision matrix

Which exact machine and stack were tested?

The operator report is dated 6 March 2026 and names an M4 Pro with ten performance and four efficiency CPU cores, twenty GPU cores and 64 GB unified memory. It uses MLX 0.31.0 and mlx-lm 0.30.7, one warmup followed by twenty measured iterations. These are historical Qwen3.5 results, not measurements of current Qwen3.8 or another Apple chip.

Source: weklund · pinned document ↗

Why are the generation and latency tables separate?

Generation uses a 256-token prompt and 100-token output. First-token latency and peak-memory tables use a 4,096-token prompt. The 4B int4 example consequently reports 68.0 ± 0.4 generated tokens/s in one profile and 6.14 ± 0.25 seconds to first token in the other. Those values cannot be assembled into one measured end-to-end 256-token request.

Source: weklund · pinned document ↗

What does the dense versus MoE example show?

The selected int4 rows put 27B at 11.8 ± 0.2 tokens/s and 35B-A3B at 25.6 ± 0.5 under the generation profile. At the longer prompt their reported peaks are 22.1 and 23.7 GiB. That is an observed speed/memory tradeoff on this stack; it does not prove equal coding quality or a broad rule that one architecture is better.

Source: weklund · pinned document ↗

Which claims are not reproduced here?

The source adds qualitative recommendations and secondary-source quality claims after its measured tables. We keep those separate from the directly reported MLX timing and memory cells. Missing BF16 cells stay missing; they are not zero-speed or failed measurements. Weight hashes, cache state and concurrent-user counts are not specified by the selected report.

Source: weklund · pinned document ↗

Reported figures and their conditions

Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.