netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit

AutoTrust-JEV-27B v1.1: MLX 4-bit

Built and tested on Apple M5 Max (128 GB) · upstream revision AutoTrust-JEV-27B v1.1 (ae15690)

A 4-bit MLX build of autotrust/AutoTrust-JEV-27B-v1.1 for Apple Silicon. 16 GB. Released 19 October 2026.

AutoTrust-JEV-27B v1.1 is a text decision model built on vllm-sr/d3 and Qwen3.8-27B. Each call is one question about a state: a choice among up to 255 options you supply, or a yes/no (noul), answered in one forward pass with a probability for every option. It classifies, routes, gates and ranks; it does not chat or generate text.

This build: Every layer at 4 bits (affine, group size 64), except the decision read-out rows in lm_head, which stay at 8 bits. Recommended for Macs with 24 to 32 GB. The same answer as the bf16 weights on 227 of 231 test decisions (KL 0.004 on average), at about half the size of the 8-bit build.

Published by NetProjects (Internet Projects Ltd). Not affiliated with AutoTrust AI.

Run it with autotrust-jev27b-mlx

The model reads its prompt bidirectionally in its full-attention layers, as it was trained (autotrust serve it on vLLM with StartLux's prefix-bidirectional plugin). Loading these weights with stock mlx-vlm or mlx-lm reads the prompt causally and gives different answers, so use autotrust-jev27b-mlx, which reproduces the attention, the model's own prompt format (jev_format.py, included here) and its read-out.

pip install "git+https://github.com/internet-projects/autotrust-jev27b-mlx"

import jev27b_mlx

jev = jev27b_mlx.load("netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit")
q = jev.question("choice", "Which team should handle this ticket?",
                 ["billing", "shipping", "technical support", "account security"])
r = jev.decide("Customer message: my card was charged twice for the same order.", q)
print(r["choice"], r["probabilities"])

Or as a server: python -m jev27b_mlx serve netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit (POST /v1/decide).

Verified against an independent reference

Measured on a MacBook Pro M5 Max with 231 decisions from the JevBench public set, rendered in the model's own prompt format. Reference: the original bf16 weights; on 39 of the prompts also an independent PyTorch implementation (transformers, attention in float32 without a mask in the full-attention layers).

build	size	memory in use	same answer as bf16 (231)	probability drift, KL mean / p95	same answer as PyTorch (39)	median per decision
8-bit	28 GB	31 GB	231 / 231	0.0002 / 0.0004	39 / 39	318 ms
4-bit	16 GB	18 GB	227 / 231	0.004 / 0.019	39 / 39	285 ms

• The MLX engine on the original bf16 weights matches the PyTorch reference on 39 of 39 decisions (KL 0.0001).
• JevBench public set: the original weights answer 209 of 231 correctly on MLX, the 8-bit build 209, the 4-bit build 210.
• Speed: one decision at a time, prompts of about 150 tokens, with other models running on the same Mac. Prompts of 3,000 to 4,000 tokens take 3.5 to 6 s. Memory in use is measured on a 3,970-token prompt.
• Text only.

For the model's own Decision Index results, see the original model card.

More from NetProjects

• autotrust-jev27b-mlx on GitHub and the NetProjects Hugging Face organization
• JEV-35B on MLX, autotrust's mixture-of-experts decision model, with vision
• H2O-Lightning-4B on MLX

Credits and license

Model: autotrust/AutoTrust-JEV-27B-v1.1 by AutoTrust AI (Apache-2.0, the same license as the original; upstream notices in NOTICE.base and LICENSE.base-apache-2.0), built on vllm-sr/d3 and Qwen/Qwen3.8-27B. Bidirectional prompt reading as specified by StartLux-Decision. Converted with mlx-vlm. Port by Paris Michalas, Partner, NetProjects (Internet Projects Ltd). Not affiliated with AutoTrust AI.

More NetProjects MLX builds: NetProjects MLX builds collection · all models at huggingface.co/netprojects