MBZUAI-Incubated Apertix Open-Sources Cerno 9B Small Decision Model

According to the report, Cerno 9B v1.2 is second among open models under 25 billion parameters on the full-score leaderboard of the Decision Index. The model ranked first is baselabs' elephant-1. Apertix said its model has a median latency of 74 milliseconds. Elephant-1 takes 114 milliseconds and the Jev reference benchmark takes 196 milliseconds, which makes Cerno about 1.54 and 2.65 times faster respectively.

The leaderboard relies on private evaluations, and 40 per cent of them test tasks in new domains. Cerno 9B scored 82.0 on tool use, compared with 74.4 for Jev. On the public version of the Decision Index it scored 66.83, up from 58.59 for the previous release, v1.1. Apertix trained the model from Qwen3.5-9B.

The family also includes 27-billion and 4-billion-parameter versions. Cerno 4B v1.2 is tied for third place with Perplexity Decider Flash v1.2 among models of up to 8 billion parameters.

What Decision Models Do

Decision models do not generate free text the way large language models do. They choose from a fixed set of answers, such as yes or no, a category, a ranking or a score. Cerno accepts text or JSON as input and returns a probability for every possible answer in a single forward pass. Apertix pitches it for the many small choices inside software products, such as routing requests, classifying inputs and picking an AI agent's next action.

Models like this are smaller and faster than general-purpose LLMs, and they are more flexible than a classifier trained for a single task. Because the output includes a confidence level, calls with low confidence can be passed to a human. Companies use them for tool selection, safety checks and support ticket triage, according to the report.

Why It Matters

As companies build agentic AI systems, every step an agent takes adds cost and delay. A fast, cheap model that runs on the device and handles routing and next-step decisions could take part of that load off larger models. Because the release is open under a permissive licence, businesses can deploy and adapt it without paying licence fees.

The release also shows MBZUAI's incubation programme producing commercial model work in Abu Dhabi alongside the university's own research. Apertix builds multimodal edge models for vision, language and audio that run on devices. Its products include Apertix-VL, the EdgeDeploy SDK and the Cascade Engine. The team says its members have more than 500 publications and more than 10 million model downloads to their name.

The benchmark figures come from the company and the Decision Index leaderboard. They have not been independently verified.