Fastino's GLiDE on a Mac: a 27B decision model that answers with probabilities, not text

Paris Michalas, Partner, NetProjects (Internet Projects Ltd)30 Oct 2026

Many questions a website asks are not open-ended. Does this customer want a refund, an exchange or information? Is this customer angry? How relevant is this page, on a scale? A model that only has to choose can give a probability for every option instead of writing a sentence, and those probabilities tell you how sure it is.

Fastino's GLiDE is built for exactly that. It does not chat or generate text. We ported it to MLX so it runs on a Mac with Apple Silicon, and published 8-bit and 4-bit builds with an engine that serves the same API as GLiDE's own server.

Always up to date. We keep these builds and the engine that runs them current on Hugging Face and GitHub. When Fastino releases a new version of GLiDE, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the README on GitHub and the model card on Hugging Face carry the current figures and a dated change log.

Hugging Face: https://huggingface.co/netprojects
GitHub: https://github.com/internet-projects
NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds

What the model does

Fastino GLiDE (v1.2) is a decision model built on Qwen3.8-27B. You give it a state, which can be text, JSON or images, and typed questions: multiple choice, yes/no, or a score. It returns a probability for every option in one pass. Inside, 16 full-attention layers read the whole prompt in both directions, alongside 48 Gated DeltaNet layers, and a 255-way readout head turns the result into answers. Fastino's original is a 52.2 GB bf16 model: https://huggingface.co/fastino/GLiDE

It sits next to the other decision models we run on our Macs: H2O-Lightning-4B, AutoTrust AI's JEV models and IBM's Granite Switch 4.1.

What we released

Measured on GLiDE v1.2, MacBook Pro M5 Max (128 GB), 29 October 2026.

Hugging Face	Size	Same top answer as Fastino's code (204 questions)	JevBench public set (231)
netprojects/GLiDE-mlx-8bit (recommended)	28.2 GB	202 of 204	215 of 231
netprojects/GLiDE-mlx-4bit	15.3 GB	197 of 204	213 of 231

Alongside the builds we released the glide-mlx engine, which serves the same API as GLiDE's own server (POST /v1/systemone). The original weights running on MLX also score 202 of 204, so the 8-bit build loses nothing measurable: against the original on MLX it gives the same top answer on 204 of 204.

How well it matches the original

We compared each build with Fastino's own code in bf16 on 72 requests holding 204 questions, 12 of them image requests.

Measured on GLiDE v1.2, 29 October 2026.

Build	Mean probability difference	Largest probability difference
Original weights on MLX	0.0016	0.025
8-bit	0.0020	0.034
4-bit	0.0115	0.124

Yes/no and score answers are identical in every build. All the differences are multiple-choice near ties: the two questions where the 8-bit build picked differently had margins of 0.018 and 0.000 between the top two options.

Speed on a Mac

Measured on GLiDE v1.2, JevBench v1.2.1 public set (231 items), MacBook Pro M5 Max (128 GB), 29 October 2026.

Build	Easy	Standard	Hard	Total	Median time per item
8-bit	48 of 48	72 of 72	95 of 111	215 of 231	0.74 s
4-bit	48 of 48	72 of 72	93 of 111	213 of 231	0.45 s

Try it

git clone https://github.com/internet-projects/glide-mlx && cd glide-mlx
pip install -r requirements.txt
hf download netprojects/GLiDE-mlx-8bit --local-dir glide-mlx-8bit
python -m glide_mlx.server --model glide-mlx-8bit

From Python:

from glide_mlx import GlideMLX
g = GlideMLX.from_pretrained("glide-mlx-8bit")
g.decide(state, questions)

A request body, from the README example:

{
  "state": "Order 123 arrived broken and the customer wants their money back.",
  "questions": {"intent": {"type": "choice", "instructions": "What does the customer want?",
                           "criteria": {"refund": "A refund", "exchange": "An exchange", "info": "Information"}},
                "angry": {"type": "noul", "instructions": "Is the customer angry?"}}}

The README has the full commands, including how to convert the weights and rerun the checks: https://github.com/internet-projects/glide-mlx

Honest limits

It must be served with glide-mlx. Generic MLX servers lack the readout head and the two-way attention, so they will not give GLiDE's answers.
The 4-bit build differs more from the original than the 8-bit build.
All figures come from one machine.
We ran only the public JevBench set, and have no JevBench score for the original weights to compare with.
There are no tokens-per-second figures, because GLiDE does not generate text; we give seconds per item instead.

Licence

Built with Fastino GLiDE.

GLiDE and our port are under the Fastino Community License. The Qwen3.8-27B base model is Apache-2.0. Commercial use must show the line Built with Fastino GLiDE, as above, and production commercial use by a company with more than USD 10 million in annual revenue needs a separate agreement with Fastino. The licence: https://github.com/internet-projects/glide-mlx/blob/main/LICENSE

Credits

GLiDE is by Fastino, built on Qwen3.8-27B by Alibaba Cloud. The MLX port is by Paris Michalas, Partner, NetProjects (Internet Projects Ltd). Not affiliated with Fastino.

Paris Michalas is a partner at NetProjects, the trading name of Internet Projects Ltd: https://netprojects.com. This article was drafted with AI assistance, then edited and approved by people on our team. You can read more about how we use AI at https://netprojects.com/standards/ai.