公開MLX変換は本当に動くか — 使えない変換を実測で見分ける方法Even models published on Hugging Face as MLX conversions can be broken — one…

Anyone assembling local large language model pipelines on Apple hardware has likely learned to treat the "MLX converted" label on Hugging Face as a signal that a model is ready to run. A recent hands-on report from the Zenn community challenges that assumption, demonstrating that some models published under this label do not work in practice — one fails to load at all, and another produces nothing but garbled characters. For practitioners who depend on community conversions rather than converting weights themselves, it is a pointed reminder that a published file is not the same as a verified, functioning model.

The case study centers on Baidu's Unlimited-OCR, a 3.3-billion-parameter optical character recognition model released under the permissive MIT license. MLX is Apple's array and machine learning framework built for Apple Silicon, and the surrounding tooling — libraries such as mlx-lm and mlx-vlm — converts models from formats like PyTorch or safetensors into a representation that runs efficiently on Mac GPUs. Because that conversion is largely mechanical, it is tempting to assume the output faithfully reproduces the original.

The author set out to convert Unlimited-OCR independently and, in the process, benchmarked two conversions that were already publicly available, running each through the same evaluation set. The results were stark. One of the two existing conversions could not be loaded at all. The other loaded but returned completely garbled output, rendering it effectively useless for the OCR task it was meant to handle. One of these flawed conversions was, notably, the most-downloaded conversion in its lineage — a detail suggesting that download counts are a poor proxy for whether a model actually works.

The report frames this as a technical observation rather than a criticism of the people who uploaded the conversions. Its central argument is that there is a real distance between "the weight files were generated" and "the model functions as a model," and that this gap can only be closed through empirical measurement. In other words, a successful export step tells you little about downstream behavior; only running the model against real inputs and inspecting the outputs can confirm usability.

There are technical reasons this class of failure is common with models like Unlimited-OCR. OCR and other vision-language systems are not monolithic — they typically combine a vision encoder, a projection layer, a language model, and a specific tokenizer or image processor. Conversion tooling tuned for text-only LLMs can mishandle these components, misalign weight tensors, or drop the preprocessing configuration a model needs. Garbled character output, in particular, frequently points to a tokenizer mismatch or an incorrect weight mapping, while a hard load failure often reflects unsupported layers or metadata that the