---
license: other
license_name: fastino-community
license_link: LICENSE
base_model: fastino/GLiDE
base_model_relation: quantized
library_name: mlx
pipeline_tag: image-text-to-text
tags: [mlx, decision-model, decision-making, calibration, jevbench, qwen3_5, apple-silicon, glide]
---
# GLiDE — MLX 4-bit

*Built and tested on Apple M5 Max (128 GB) · upstream revision GLiDE v1.2 (`d199b6d`)*

A 4-bit MLX conversion (affine, group size 64) of [fastino/GLiDE](https://huggingface.co/fastino/GLiDE) (GLiDE v1.2, revision `d199b6d`), Fastino's decision model built on Qwen3.8-27B, for Apple Silicon. 15.3 GB of weights.

**Built with Fastino GLiDE.** Published by [NetProjects](https://netprojects.com) (Internet Projects Ltd). Not affiliated with Fastino.

GLiDE is a **decision model**, not a chat model. Give it a state (text, JSON or images) and typed questions (`choice`, `noul` for yes / no, `score`) and it returns a probability for every option, in one forward pass per question. Nothing is generated.

> **Serve it with [glide-mlx](https://github.com/internet-projects/glide-mlx), not with `mlx_lm.server` or a plain `mlx_vlm` generate call.** GLiDE reads its answer from a separate 255-way readout (`readout.safetensors`) and lets its full-attention layers read the whole prompt in both directions; generic MLX snippets do neither.

## Results (MacBook Pro M5 Max, 128 GB, 29 October 2026)

| | this checkpoint |
|---|---|
| Same top answer as Fastino's own PyTorch code (bf16) | **197 / 204** questions; probability difference mean 0.012 |
| Yes / no and score questions | 132 / 132 identical |
| JevBench public (231, JevBench v1.2.1, `--adapter typesafe`) | **213 / 231** (easy 48/48 · standard 72/72 · hard 93/111) |
| Time per JevBench item (prompts up to about 4K tokens) | 0.45 s median |

The answers that differ from Fastino's code are all `choice` questions with many options where the reference's top two options were close (top-two margin 0.09 or less). The comparison set has 72 requests with 204 questions: text and JSON states of 30 to 1,500 words, 12 image requests, and all three question types. The smallest build: about half the size of the 8-bit build, with probabilities that move a little more. For the closest match to the original, use [netprojects/GLiDE-mlx-8bit](https://huggingface.co/netprojects/GLiDE-mlx-8bit).

## What is kept as in the original

- **Reading:** the 16 full-attention layers see the whole prompt (no causal mask), the 48 Gated DeltaNet layers keep their recurrence, as GLiDE's `noncausal_full_attention`.
- **Answer:** the last token's hidden state goes through GLiDE's 255-way readout, kept at its original precision; softmax over the options at the temperature in `decision_config.json`.
- **Prompts and images:** GLiDE's own `prompt.py` and `images.py`, unchanged; the vision encoder stays at its original precision.
- **API:** the same `POST /v1/systemone` request and response format as GLiDE's server.

## Use

```bash
git clone https://github.com/internet-projects/glide-mlx && cd glide-mlx
uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -r requirements.txt
hf download netprojects/GLiDE-mlx-4bit --local-dir glide-mlx-4bit
.venv/bin/python -m glide_mlx.server --model glide-mlx-4bit --port 8790    # http://127.0.0.1:8790/v1/systemone
```

Python: `from glide_mlx import GlideMLX; g = GlideMLX.from_pretrained("glide-mlx-4bit"); g.decide(state, questions)`.

## License

GLiDE is licensed under the **Fastino Community License** (`LICENSE`, `NOTICE`, `licenses/Qwen3.8-27B-LICENSE` are included). This conversion is a Derivative Work under that license and is distributed under the same terms. In short: commercial use requires displaying "Built with Fastino GLiDE" (section 4), and production commercial use by an entity with annual revenue above USD 10 million requires a separate agreement with Fastino (sections 2–3). Read the license for the full terms.

Credits: Fastino (GLiDE), Qwen team (Qwen3.8-27B, Apache-2.0), MLX and mlx-vlm. Conversion by Paris Michalas, Partner at [NetProjects](https://netprojects.com). Not affiliated with Fastino.

**More NetProjects MLX builds:** [NetProjects MLX builds collection](https://huggingface.co/collections/netprojects/netprojects-mlx-builds-6ae3b2683a59a1b4fc4c0c3a) · [all models at huggingface.co/netprojects](https://huggingface.co/netprojects)