Skip to content
NetProjects
Menu

Insights

Two AutoTrust AI decision models on a Mac: JEV-35B and JEV-27B on MLX

Paris Michalas, Partner, NetProjects (Internet Projects Ltd)

Some questions don't need a chatbot. Which team should handle this ticket? Is this photo of a damaged box? Is this search result about what the person asked? A decision model answers questions like these in one step and gives a probability for every option, so you can see how sure it is.

AutoTrust AI publishes two such models under the JEV name. On 19 October 2026 we released MLX builds of both, so they run on a Mac with Apple Silicon: no GPU server and no per-call API bill.

What we built

JEV-35B is a mixture-of-experts model (35 B parameters, about 3 B active per token) that can also look at images. It answers three kinds of question: a choice among up to 256 options, a yes/no, or a score from 0 to 5. We converted it with its vision tower intact and released three builds:

  • 8-bit, 36 GB, for Macs with 48 GB or more
  • 4-bit mixed, 21 GB, for Macs with 32 GB
  • 3-bit mixed, 17 GB, for Macs with 24 GB

In the mixed builds only the experts are squeezed to 4 or 3 bits. Every other layer stays at 8 bits, and the vision tower and the router gates that pick the experts stay at full bf16 precision. That choice matters: keeping those layers precise made the 4-bit build about six times closer to the original than a plain 4-bit conversion, for one extra gigabyte.

AutoTrust-JEV-27B v1.1 is a dense text model, released by AutoTrust AI on 17 October. It was trained to read the whole prompt in both directions before answering, rather than strictly left to right. The standard Mac tools, mlx-vlm and mlx-lm, read prompts left to right only, so they load the weights without error and quietly give different answers: in our test, 5 of 39 answers changed. We wrote a dedicated engine, jev27b-mlx, that reads the prompt the way the model was trained, and released two builds:

  • 8-bit, 28 GB, for Macs with 48 GB or more
  • 4-bit, 16 GB, for Macs with 24 to 32 GB

Does the port match the original?

We checked every build against the original weights on 231 decisions from the JevBench public set.

  • JEV-35B: the 8-bit build gives the same answer as the original on 228 of 231, the 4-bit mixed build on 225, the 3-bit mixed build on 224. Every build also passes our image check, 4 out of 4.
  • JEV-27B v1.1: the 8-bit build gives the same answer as the original on all 231 decisions, and the 4-bit build on 227. We also checked our engine against an independent PyTorch implementation, which it matches on 39 of 39 decisions. On the JevBench public set the original weights answer 209 of 231 correctly on the Mac, the 8-bit build 209 and the 4-bit build 210.

Answers are quick. On a MacBook Pro M5 Max, with short prompts, a typical JEV-35B decision takes about 100 to 165 ms, and a JEV-27B decision about 0.3 s.

Putting them to work: ranking search results

We run our own search gateway for our AI tools. A small dedicated reranker (Qwen3-Reranker 0.6B) sorts the results, and we wanted to know whether the JEV models could do better.

We took 100 real searches from our logs, fetched 12 results for each (1,200 in all) and had an independent judge grade every result. Then we ranked the results four ways. For the JEV models (JEV-35B 4-bit mixed, JEV-27B 4-bit) we used AutoTrust AI's own recipe: one yes/no question ("is this result about what the search asks for?") and one 0 to 5 usefulness score per result, averaged.

  • JEV-27B ranked best. It put a fully relevant result first in 71% of searches, against 59% for the dedicated reranker. On the standard ranking measure (nDCG@6) it scored 0.85, against 0.81 for the reranker and 0.61 for the search engine's own order.
  • JEV-35B matched the dedicated reranker (0.80) and put a fully relevant result first slightly more often (64% against 59%).

The dedicated reranker remains the fast choice for live search, at about half a second per search against about 8 seconds for JEV-27B. Where quality matters more than a few seconds, such as research for a long article, JEV-27B gives the better ranking.

Try it

Install the engine for the model you want and point it at our builds on Hugging Face:

JEV-35B, then load netprojects/JEV-35B-mlx-4bit-mixed:

pip install "git+https://github.com/internet-projects/jev35b-mlx"

JEV-27B v1.1, then load netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit:

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

Both engines also run as a small local server with a /v1/decide endpoint; the JEV-35B one uses the same request fields as AutoTrust AI's own.

What we are releasing

All five builds, both engines, the conversion scripts and the scripts that reproduce the checks against the original weights, under the Apache-2.0 licence, the same licence as the original models.

Credits

The models are by AutoTrust AI. JEV-35B is a fine-tune of Qwen3.5-35B-A3B; AutoTrust-JEV-27B v1.1 is built on vllm-sr/d3 and Qwen3.8-27B, and the two-way prompt reading follows StartLux-Decision. The builds were converted with mlx-vlm. NetProjects is not affiliated with AutoTrust AI.

Links

Paris Michalas is a partner at NetProjects, the trading name of Internet Projects Ltd. 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.