#UAE #AImodels - Edge AI start-up Apertix, from the incubator of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), has open-sourced the latest in its Cerno family of small decision models: Cerno 9B v1.2. The nine billion parameter model ranks second among open models under 25 billion parameters on the Decision Index leaderboard (full score), behind only baselabs’ elephant-1. Apertix says the new model’s median latency is 74 ms, against 114 ms for elephant-1. The new model is available via Hugging Face on an Apache 2.0 licence.
SO WHAT? - Most AI models are focused on writing text, but Cerno is built to handle agentic AI tasks running on edge devices. The new small decision model reads a situation and returns probabilities for a set of choices in one pass. This makes it ideal for tasks like routing requests or deciding an AI agent’s next step. The model has been developed to be small, fast and cheap to run and is now available as open-source.
KEY POINTS:
Abu Dhabi-based edge AI startup Apertix has released Cerno 9B v1.2 as open source. The Cerno family includes 27B, 9B and 4B models, and the weights and inference code are under an Apache 2.0 licence.
Cerno 9B ranks second among open models under 25 billion parameters on the Decision Index Full Score leaderboard. Only baselabs’ elephant-1 ranks higher. The board uses private evaluations, 40% of them on new-domain tasks.
According to the Decision Index the new model has a median latency of 74 ms, which is 1.54 times faster than elephant-1 at 114 ms, and 2.65 times faster than the reference Jev benchmark at 196 ms.
Cerno 9B scores 82.0 on tool use, ahead of Jev’s 74.4. It is 4.7 points clear of the next 9B entry, vLLM-SR d3-flash, which scores 54.3.
Cerno 4B v1.2 is tied in third place with Perplexity Decider Flash v1.2 in the up-to-8 billion parameters bracket.
Cerno is built for the many small decisions inside product workflows, such as routing requests, classifying inputs and choosing an AI agent’s next action. It takes text or JSON and returns a probability for every answer option in one forward pass.
The Cerno 9B v1.2 model is trained from Qwen3.5-9B and scores 66.83 on the public Decision Index, up from 58.59 for v1.1.
Apertix builds multimodal edge AI models for vision, language and audio that run on-device. Its products include Apertix-VL, the EdgeDeploy SDK and the Cascade Engine. Its team says its research record spans 500-plus publications and 10 million-plus model downloads.
Apertix is participating in the AI incubator at the MBZUAI Incubation and Entrepreneurship Center (IEC).
ZOOM OUT - Small decision models are compact AI models that choose an answer from a fixed set of options, such as yes or no, a category, a ranking or a score with a confidence level. These models don’t generate text the way a large language model (LLM) does. That makes them faster than an LLM and more flexible than a classifier trained for one task, meanwhile low-confidence calls can be routed to a human. Companies use small decision models to route requests, select tools, run safety checks and triage support tickets.
[Written and edited with the assistance of AI]
Source: Apertix
LINKS
Cerno family model page (Hugging Face)
Cerno model code (GitHub)
Decision Index Leaderboard (Hugging Face)
Read more about AI models:
MBZUAI builds video AI that plans its own next steps (Middle East AI News)
Egyptian startup releases three open-source Arabic models (Middle East AI News)
MBZUAI unveils AI for autonomous biomedical discovery (Middle East AI News)
New Arabic AI Atlas maps 1,000+ open models and tools (Middle East AI News)
TokenAI launches compact AI decision model (Middle East AI News)