There has been some excitement over decision models in the AI world lately. A decision model is like an LLM but it does not generate text in its output. It accepts a prompt, but returns JSON output and a list of probabilities—so basically you pass it a question and it responds with the answer (or likelihood of the answers).
Decision models are on the systemone end point. I downloaded Kev-4B-Q4_K_M.gguf, which is an open-weights version of Jev called Kev. It is based on Qwen 3.5 4B, and is only 3 GB. I then set up Llama to use this model.
With curl commands, I can send questions and instructions to the model. For example, I used the question "what color is the sky?" and offered the criteria blue, green, red, black, yellow.
When I used the prompt "the sky is azure", Kev responded with "blue" and "green" with probabilities 0.96 and 0.02; the other colors were all 0.006 or below. So it understood "azure" means blue and green. Then I tried "it is night time in the countryside." Kev responded with the probabilities black (0.82) and blue (0.08) and red (0.06). The model linked "night time" with a dark sky, possibly with blue or red (due to the sunset). It responded instantly in both tests; Kev is a promising way to classify or decide based on prompts.