How it compares? I’ll bet this one will be definitely faster than nvfp4 version just by looking at the size. “Intelligence” for sure is going to be different and depending on the task you might notice or not the lacking bits of depth.

According to my LLM expert evaluator (GLM-5.1), after 20 tests of coding and tool calling that would take me weeks or months to solve it gave 3% higher score to vanilla 4-bit mlx 😑 (whatever that means, something like 3% less errors than 4-bit, but also different errors that those committed by 4-bit model) … but did I notice a difference or does it even matter for my use case? I then gave both models a testing set of multi step multi skill tasks based on actual things I need to do daily on Hermes agent and both completed it, so it was a tie FOR MY USE CASE, ymmv kind of thing.

Nothing fancy really, a simple 2bits of difference between embedding and prediction layers against the rest, but instead of the more conventional 4&6 or 6&8 mixes I went really hard to 3&5… no datasets or nothing to be “aware” during process, uniform within layers and differentiated between layers groups.

Qwen3.6 model is so robust and dense that it still holds together. Not all models are this resilient, there was a post here some days ago comparing divergence from bf16-fp8-q4 between Gemma 31b and Qwen3.6… Qwen hold it together much better.