Introducing Deca 2.5 Ultra: Now in Beta!

Unlock unparalleled intelligence with our most advanced model yet.

We’re releasing Deca 2.5 Ultra in a private beta testing group, a high-performance AI model built for complex tasks requiring substantial computational power. Ultra is built on the DynaMoE architecture, which uses expert routing mechanisms to allocate resources dynamically, optimizing performance across a variety of tasks. The model incorporates DynaMoE's routing to a Vision-Language Model (VLM) cross-attention mechanism, enabling advanced image processing and multi-modal reasoning. This allows Ultra to process textual and visual data in a highly integrated way, making it especially adept at tasks that require both linguistic and visual context. Ultra excels in long-context reasoning, with specially trained experts dedicated to handling long-range dependencies. These experts allow the model to maintain accuracy and coherence over extended inputs, which is critical for tasks that require deep understanding across large spans of text. In addition, Deca 2.5 Ultra has a very strong spatial sense, thanks to a set of experts specifically trained in spatial reasoning and design tasks. This enables the model to perform complex reasoning related to spatial relationships, layout, and design, making it well-suited for tasks in fields like architecture, computer-aided design (CAD), and other spatially-intensive domains. The model also uses parallel reasoning, meaning it processes multiple aspects of a task simultaneously rather than relying on streaming. This design provides faster, more efficient inference, especially for tasks requiring multi-step reasoning and decision-making. We’re inviting developers and researchers to join the beta program to help refine Ultra’s performance, stability, and reliability. Your feedback will be crucial in optimizing the model for a broader release and ensuring it meets the highest standards of AI performance. Get Started with Deca 2.5 Ultra Beta →