Skip to content

v0.1.0 — GLiDE v1.2 on MLX

Latest
Latest

Choose a tag to compare

View all tags
@pmichalas pmichalas released this 29 Oct 10:19
· 1 commit to main since this release
v0.1.0
8f641ad

GLiDE v1.2 on Apple Silicon (MLX). First release of the MLX port of fastino/GLiDE (revision d199b6d), Fastino's calibrated decision model built on Qwen3.8-27B. Built with Fastino GLiDE. Not affiliated with Fastino.

Builds

  • netprojects/GLiDE-mlx-8bit, 28.2 GB, recommended: the same top answer as Fastino's own PyTorch code on 202 / 204 test questions (204 / 204 against the original weights on MLX); JevBench public 215 / 231.
  • netprojects/GLiDE-mlx-4bit, 15.3 GB: 197 / 204; JevBench public 213 / 231.

What's in the package

  • glide_mlx: loader, decide() / decide_batch(), and a server with GLiDE's own POST /v1/systemone API.
  • Same reading as the original: the full-attention layers see the whole prompt, the Gated DeltaNet layers keep their recurrence, and the answer comes from GLiDE's 255-way readout. GLiDE's prompt and image code is included unchanged.
  • convert.py (8- and 4-bit), parity.py (side-by-side check against Fastino's code), speed.py, and a tiny-model test that checks the port without the 52 GB download.
  • Steady memory: the MLX buffer cache is capped and cleared after each request.

Measured on a MacBook Pro M5 Max (128 GB), 29 October 2026. License: Fastino Community License (Qwen3.8-27B base: Apache-2.0). Port by Paris Michalas, Partner at NetProjects.

Assets 2
Loading