What happened to oQe?

[deepsweet]

v0.3.1 changelog:

oQ: update descriptions to reflect current implementation, temporarily disable enhanced quantization UI

127b08e commit:

refactor: clean up oQ codebase for upcoming enhanced quantization redesign

Are oQe quantizations already uploaded to HuggingFace considered obsolete?

[jundot]

Thanks for raising this, and thanks for all the oQ test data you've been sharing. It's been really useful input.

Short answer: the oQe quants already on HuggingFace aren't broken. They still load and run fine. But I'm not going to be re-uploading or refreshing them while the enhanced path is paused, so consider them frozen rather than actively maintained.

On why it's paused: in the few weeks before I disabled it, I tried several variants of the enhanced approach. GPTQ-style, AWQ-style, the original formulations, plus a few MoE-optimized variations of each. Across a handful of models and short benchmarks, the results were not consistent. Some models clearly improved with the enhanced path. Others actually regressed compared to the plain oQ baseline. I haven't found a single method that reliably improves quality across model architectures, so I'd rather keep it off in the UI than ship something that silently makes some models worse. Once I find an approach that holds up across architectures, I'll bring enhanced quantization back.

On benchmarking more broadly: from my own testing, MMLU alone isn't enough to judge quantization quality. A single benchmark can flatter or punish a method depending on the model, so proving it properly means running diverse benchmarks across many models, and that takes a lot of time. As a solo developer I have to prioritize where my time goes, and right now bug fixes, new features, and new model support take precedence. I'd love to keep digging into oQe, but I just can't put as much time into it as I'd like to right now. Thanks for understanding.

[nickcolea]

Thanks for your honest answer.

In the meantime, do you have any suggestion about Qwen3.6 variants to use (oQ4 or oQ4e) for example?

You mentioned some models regresses, so naturally I'm curious if Qwen3.6 was suffering as it's my daily driver for now.

[deepsweet]

@nickcolea let's add it to mlx-eval results and see. What's the source of your Qwen3.6 oQ4e?

[nickcolea]

splats/Qwen3.6-27B-oQ4e and Jundot/Qwen3.6-27B-oQ4

Tried both, but don't know which would be recommended. Would be interesting to see the results you're working on 👍

[crazyi]

@nickcolea hi, can you providing some benchmark performance for the splats/Qwen3.6-27B-oQ4e and Jundot/Qwen3.6-27B-oQ4 models？