Abu Dhabi-based artificial intelligence startup Apertix has released Cerno 9B v1.2, an open-source AI model designed to make decisions faster without relying on the lengthy text-generation processes associated with conventional large language models.
The company, which participates in the incubator at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), is focusing on a less visible but increasingly important part of artificial intelligence: helping AI systems decide what to do next.
Cerno 9B v1.2 contains nine billion parameters and has secured second place among open models with fewer than 25 billion parameters on the Decision Index Full Score leaderboard, according to results reported on November 4, 2026.
The model recorded a median response latency of just 74 milliseconds, compared with 114 milliseconds for the higher-ranked elephant-1 model from baselabs.
For developers building autonomous AI agents, that difference could matter. An agent might need to make dozens of decisions while completing a single task, and delays can accumulate quickly.
Apertix Cerno 9B Focuses on Decisions Rather Than Conversations
Most attention in generative AI goes toward models that write articles, answer questions or generate software code. Apertix is taking a different route by developing compact AI systems that focus on selecting actions rather than producing lengthy responses.
Its Cerno family consists of specialized decision models designed to evaluate information and choose an appropriate response from a predefined set of options. Instead of generating paragraphs of text, the models process natural language or structured JSON inputs and return probability scores for possible decisions in a single forward pass.
Consider an AI-powered customer service platform receiving thousands of requests. One message might require a refund, another technical assistance, while a third needs immediate human attention. A decision model can determine where each request should go without asking a larger generative AI system to reason through every routine interaction.
The same approach can support automated workflows, AI tool selection and systems that need to make repeated decisions with minimal delay. Cerno is designed to handle these tasks while keeping computational demands relatively low, making it potentially useful for businesses that need fast, predictable AI responses.
Cerno 9B Ranks Among the Leading Open AI Decision Models
The latest release puts Apertix in a competitive position among smaller open-source AI models. On the Decision Index Full Score leaderboard, Cerno 9B v1.2 ranks second in the category covering models below 25 billion parameters, with only baselabs’ elephant-1 ranking higher.
Speed is one of its strongest reported advantages. Cerno 9B achieved a median latency of 74 milliseconds, making it approximately 1.54 times faster than elephant-1 and 2.65 times faster than the 196-millisecond Jev reference benchmark. The model also recorded a tool-use score of 82.0, compared with 74.4 for Jev.
Apertix has improved performance since the previous release. Cerno 9B’s public Decision Index score increased from 58.59 in version 1.1 to 66.83 in version 1.2, indicating progress in the model’s ability to handle decision-oriented tasks.
The evaluation framework includes private tests and assessments involving unfamiliar task domains. These results offer a useful comparison of model capabilities, although performance in real-world applications will depend on the hardware, workloads and deployment conditions involved.
Apertix also has a smaller model, Cerno 4B v1.2, which shares third place with Perplexity Decider Flash v1.2 in the category for models containing up to eight billion parameters. The result suggests that the company is pursuing competitive performance across multiple model sizes rather than concentrating on a single flagship release.
Open-Source Cerno Models Give Developers Greater Deployment Flexibility
Apertix has made its Cerno model family available through Hugging Face, with inference code published on GitHub. The release uses the Apache 2.0 license, allowing developers to use, modify and distribute the software under its licensing conditions.
The Cerno family includes 27-billion, 9-billion and 4-billion-parameter models, giving developers several options depending on the complexity of their applications and the computing resources available. Cerno 9B v1.2 builds on Alibaba’s Qwen3.5-9B model, with modifications intended to improve decision-making performance.
The publicly available repository provides implementation details, evaluation examples and deployment instructions. Developers can examine how the models process inputs and generate probability-based outputs instead of relying entirely on an inaccessible proprietary decision system.
For companies developing AI-powered applications, that level of control can be valuable. Organizations may want to customize how decisions are made, integrate models with existing software or operate AI services within their own infrastructure.
Open-source availability also allows independent researchers to examine the model’s performance and identify areas where improvements are needed. That does not guarantee widespread adoption, but it gives Cerno an opportunity to reach developers who prefer flexible deployment options over closed AI services.
Why Smaller AI Models Matter for Agentic AI
Agentic AI systems are becoming more capable of carrying out tasks with limited human involvement. Yet completing a single assignment often requires numerous smaller decisions that do not necessarily need the capabilities of a powerful generative AI model.
An autonomous assistant might need to identify the right software tool, determine whether a request requires approval or decide when a task should be transferred to another system. These actions can occur repeatedly throughout an automated workflow.
Using a large language model for every decision can introduce additional processing costs and delays, particularly when applications need to handle high volumes of requests.
Smaller decision models offer another approach. They can manage specific choices while leaving complex reasoning, writing and problem-solving to larger AI systems. This division of responsibilities could help developers build AI agents that operate more efficiently without sacrificing the capabilities required for complicated tasks.
Businesses may find the approach useful in customer support, IT operations, financial services and internal workflow management, where systems frequently need to classify information or select the next action.
Reliability remains important. A fast decision model is not automatically an accurate one, especially when handling unfamiliar inputs or situations involving sensitive information. Developers still need to evaluate how models behave under different conditions and establish safeguards for decisions that require human oversight.
MBZUAI Incubator Supports Abu Dhabi’s Growing Edge AI Ecosystem
Apertix’s development reflects the expanding role of Abu Dhabi in artificial intelligence research and commercialization. The startup participates in the incubation program at MBZUAI’s Incubation and Entrepreneurship Center, which supports emerging AI businesses as they develop technologies for commercial applications.
Beyond Cerno, Apertix is developing multimodal edge AI technologies capable of processing visual, language and audio information. Its portfolio includes Apertix-VL, the EdgeDeploy SDK and the Cascade Engine, technologies aimed at bringing AI capabilities closer to the devices and systems where data is generated.
According to the company’s reported figures, its team’s research background includes more than 500 publications and over 10 million model downloads. These figures highlight the technical experience associated with the startup, although they should not be interpreted as adoption figures for Cerno itself.
Edge AI is attracting attention because it can reduce dependence on centralized cloud infrastructure. Processing information locally may improve responsiveness, lower communication delays and help organizations manage sensitive information, depending on how their systems are designed.
For the UAE, startups developing these technologies contribute to a broader effort to strengthen domestic AI expertise and turn research into commercially useful products. MBZUAI’s involvement in supporting emerging companies also creates opportunities for academic research to move beyond laboratories and into practical applications.
Cerno 9B Highlights a Different Direction for AI Development
The release of Cerno 9B v1.2 comes as developers increasingly look beyond model size as the main measure of artificial intelligence progress. While large foundation models continue to dominate discussions around advanced reasoning and generative capabilities, many practical applications depend on much simpler tasks.
Recognizing intent, choosing an action, assigning priorities or determining when human intervention is necessary may not require an enormous AI model. What matters in these situations is whether the system can make the right decision quickly and consistently.
Apertix is betting that specialized models can handle this work more efficiently than general-purpose AI systems. Its latest benchmark results suggest that compact decision models can deliver competitive performance without requiring the largest parameter counts.
The bigger question is how well Cerno performs outside controlled evaluations. Production environments introduce unpredictable requests, integration challenges and operational requirements that benchmarks cannot fully capture. Adoption will ultimately depend on accuracy, reliability, deployment costs and the ease with which developers can incorporate the models into existing systems.
Still, Cerno 9B represents an interesting development for the UAE’s AI ecosystem. Rather than competing solely to build larger conversational models, Apertix is targeting the decision-making processes that could become increasingly important as autonomous AI agents move into everyday business operations.
For developers, the open-source release provides another option for building AI systems that do more than generate responses. They need to make decisions, too.
