GitHub radar
The AI model that scores probabilities, not paragraphs
GEV-26B-Decide answers a multiple-choice question with a calibrated probability for each option, instead of a paragraph of text. It's built for repeated decisions, like lead scoring or ticket routing.
The model takes a question with answer options and scores a probability for each one, instead of writing text. Calibrated means the probability matches how often that answer turns out to be correct. You can turn on a thinking mode: then the model answers fast first, and if it isn't confident, it reasons step by step. It also works with images. It can click the right spot on a screen, or move a robot arm based on a camera photo. Among its related models, it's the fastest at clicking — three times faster than JEV-27B-VL. It's worse at actually grasping objects: 40% success versus 75% for that same model. There's no ready-made package for Ollama. Ollama is a tool that runs models locally with one click. To run GEV-26B-Decide you need vLLM or transformers instead — these are developer tools that run AI models. You'll need a Mac with at least 32 GB of memory, or a GPU with 24 GB.
When it helps
Useful if you need to classify or score a stream of similar decisions — tickets, leads, reviews. Not a fit if you want a chat assistant that writes connected text.
Pros
- AutoTrust's own benchmark beats paid competitor TypeSafe Jev
- Thinking mode sharply raises accuracy on logic puzzles
- Fastest model in its family at screen clicks
- Handles both text and image inputs
Cons
- Competitor comparison is AutoTrust's own benchmark, not independent
- Worse at grasping objects with the robot arm: 40% vs 75%
- No Ollama package; you need vLLM or transformers instead
How to set it up — step by step
- 1Open the model page on Hugging Face — a site where AI models are published: https://huggingface.co/autotrust/GEV-26B-Decide
- 2Check you have a Mac with at least 32 GB of memory, or a GPU with 24 GB.
- 3Open your AI agent (Claude Code, Codex) inside your project folder.
- 4Send the agent the text from the block below.
- 5Check the agent ran the model via vLLM or transformers and got probabilities back.
- 6Ask the model your own typical question with answer options and compare the probabilities.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Figure out how to run AutoTrust's GEV-26B-Decide model locally from https://huggingface.co/autotrust/GEV-26B-Decide via vLLM or transformers, ask it a multiple-choice question, and show me the probability it returns for each option.
You need a Mac with 32 GB of memory or more, or a GPU with 24 GB of VRAM — the model has 26 billion parameters (about 4 billion active per token). There is no ready-made Ollama build.
Open on Hugging Face▌ More finds