ouuaagghh aahh ahhgg I'm getting poaached. labs! help! they're poaching meee

MOSS-Image itself is gorgeous. Go try it out and give @Open_MOSS a follow! Demo: huggingface.co/spaces/OpenMO… Site: openmoss.ai/MOSS-Image/ Code: github.com/OpenMOSS/MOSS-Ima… ImageWeights: huggingface.co/OpenMOSS-Team… On a Mac: pip install moss-image-mlx MLX weights: huggingface.co/collections/V… My code: github.com/felix-ab/moss-ima…

It reproduces OpenMOSS’s official samples, given the same seeds. The sampler is identical to upstream, seed noise matches PyTorch’s CPU generator, and FP32 outputs match the official PyTorch pipeline to 69.7 dB PSNR! Plug in a published gallery prompt and seed and you get the same picture, locally on your Mac. mlx-vlm draws different noise, so it simply can’t.

There is zero text loss at 8-bit. Across 14 test images (signs, posters, a typed paragraph, Chinese notices, JSON layouts), the 8-bit tier matched BF16 on every prompt, and compact still nailed every line. A 4-bit text encoder scrambled a paragraph, so it didn't make the cut.

Peak memory decides what your Mac can run, so a port shouldn't balloon it. Running mine at 8-bit takes just 9.3 GB for a 1024² image (or 7.4 GB in compact mode). By contrast, mlx-vlm spikes to 38.7 GB in its final four seconds. It runs an untiled FP32 decode while keeping the 10 GB DiT in memory, all on an uncapped MLX cache. Mine avoids the spike entirely by decoding in strips.

MOSS-Image is the beautiful open image model that writes long English and Chinese text Introducing the best port to MLX. Mine vs mlx-vlm’s: – 4x lower peak memory (9.3 vs 38.7 GB) – half the download – zero text loss at 8-bit – same seed makes same pic as OpenMOSS samples 🧵

ZIT is almost a year older than Flux 3 Klein, runs in half the steps, and totally decimates it on photorealism. What other checkpoint in this size/speed bracket is even competing right now?

“Hey grok, am I the first to tweet this banger or did someone beat me to it?”

eat fruit + cheese

facebook marketplace is calling… few will answer

starting another random project