Qwen3-Coder-Next quants trade blows on Mac

The big takeaway is that Qwen3-Coder-Next looks pretty quantization-tolerant on coding tasks: the local 3-bit and 4-bit runs stayed close enough to bf16 that one-off benchmark noise could easily reshuffle the order.

• –bf16 led at 65.0% average pass rate, but the best local quants landed within a few points, which is a small gap for a single-run eval
• –MLX 4-bit slightly outperformed the GGUFs on this run, but the spread is narrow enough that it’s better read as “rough parity” than a decisive win
• –The author’s claim that MLX is not meaningfully faster than GGUFs is supported by the numbers here, especially once you factor in the reported oMLX throughput bug
• –For Mac users, this points to a practical decision tree: use whichever runtime is most stable and easiest to serve, because quality differences at 3-4 bits appear modest
• –The benchmark is still anecdotal, so it’s more useful as a sanity check than a final verdict on MLX vs llama.cpp