ChengyiX's picture
Publish Fable native-MLX migration study (paper, evidence, code)
3397fcb verified
Raw History Blame Contribute Delete
1.9 kB
cff-version: 1.2.0
message: "If you use this report or its evidence, cite it as below."
title: "From GGUF to Native MLX: A Reproducible Fable Migration and the Memory Policy That Blocked It"
version: 1.0.0
date-released: 2026-10-25
license: MIT
type: report
abstract: >-
A deployment engineering report on migrating the Qwen3.8-27B TURBO-Fable
checkpoint from GGUF to a native MLX artifact on an Apple M4 Max with 64 GiB
unified memory. It documents a bounded CPU-only conversion with full layout
verification, four interleaved speculative-decoding conditions on an
8,159-token prompt, two timed-out real agent repair sessions, and a
serving-runtime system-memory guard whose documented switch converted a
refused ~25k-token prefill into a completed reply. No 20 tok/s floor, ranked
intelligence claim, peer review or near-limit context qualification is made.
authors:
- family-names: Xu
given-names: Chengyi
- name: "Nexora"
website: "https://hiklik.ai"
keywords:
- Apple Silicon
- MLX
- MTPLX
- multi-token prediction
- DFlash2
- quantization
- local deployment
- memory policy
preferred-citation:
type: report
title: "From GGUF to Native MLX: A Reproducible Fable Migration and the Memory Policy That Blocked It"
authors:
- family-names: Xu
given-names: Chengyi
- name: "Nexora"
year: 2026
publisher:
name: "Nexora Research"
references:
- type: software
title: "MTPLX: native MTP speculative decoding on Apple Silicon"
authors:
- family-names: Altoukhi
given-names: Youssof
version: "2.12.2"
repository-code: "https://github.com/youssofal/MTPLX"
- type: software
title: "mlx-dspark: DSpark and DFlash speculative decoding for MLX"
authors:
- family-names: Rahim
given-names: A.
version: "0.19.0 and 0.20.3"
repository-code: "https://github.com/ARahim3/mlx-dspark"