Re-download this Qwen3.8 MLX quant, its maintainer says
Jundot's October 29 notice replaces an existing Flash-Next oQ4e MTP checkpoint after an oMLX calibration fix; the card lists 100.275 GiB of tensor storage.
If you already downloaded Jundot’s Qwen3.8-Flash-Next-oQ4e-mtp, its maintainer wants you to fetch it again. An explicit October 29 notice says the weights were re-quantized with improved oQe quantization in oMLX 0.7.3. This is a replacement of a community checkpoint whose Hugging Face repository dates to August 27, not a new Qwen base-model launch.
The card links the change to oMLX pull request 4930, merged on October 29 at 03:28 UTC. The maintainer describes both revised bit allocation and a fix to the calibration walk: some decoder blocks had received different inputs from those used in the model’s normal forward pass. The fix corrects how importance and sensitivity statistics are collected, and the PR says old calibration caches are not reused.
A complete MLX build, with a substantial download
This checkpoint includes the language backbone, n-gram embedding, vision encoder and multi-token prediction head. The author’s storage table puts the complete package at 107.670 GB, or 100.275 GiB, averaging 4.785 bits per weight across the checkpoint. The language backbone alone averages 4.661 bits. Those are tensor-storage figures, not a guarantee that a machine with exactly that much memory can run every context length.
It is an MLX build for the Apple Silicon inference ecosystem, rather than a GGUF download for llama.cpp. The model card points to oMLX and identifies Qwen Community License 1.0. Users should check the runner’s compatibility and available memory before replacing a working local setup; the card does not provide a universal minimum-memory or speed guarantee.
The PR reports fidelity measurements on several models, but its dense-model gains should not be copied onto this Flash-Next checkpoint. Its separate large-MoE comparison concerns the oQ3e build and uses a high-bit reference, not an fp32 reference for this exact oQ4e MTP package. The actionable news here is the maintainer’s dated replacement notice and documented calibration correction. AITrending has reviewed those English sources without independently benchmarking the new weights.