MLX NVFP4 trails 4-bit in tests

Promising support, underwhelming payoff so far. My read is that MLX has the format wired up, but the practical gains are still muted unless the kernels and native paths mature.

• –MLX v0.30.3 says it supports NVFP4 and MXFP8 quantized ops on Metal and NVFP4/MXFP8 quantized-quantized matmul on CUDA, so the framework is actively moving here: https://github.com/ml-explore/mlx/releases
• –The Reddit benchmark on Qwen 3.5 35B shows NVFP4 at 7.991 word perplexity versus 7.850 for 4-bit, with MXFP4 even worse at 8.379: https://www.reddit.com/r/LocalLLaMA/comments/1ry5gm3/has_anyone_tried_nvfp4_on_mlx/
• –Reported speed was roughly the same as 4-bit, so there is no obvious throughput upside yet on the Mac path.
• –The thread’s guess that MLX may be dequantizing or otherwise emulating the format is plausible, but that part is still an inference rather than confirmed behavior.