reka-core-20240904 is a specific production update of Reka Core, a frontier-class multimodal language model developed by Reka AI. The model is designed to natively process and reason across multiple modalities, including text, images, video, and audio. Unlike models that append vision capabilities to a text-only core, Reka Core was trained from scratch to handle diverse inputs within a single unified architecture.
The model features a 128,000-token context window, enabling it to analyze long documents and extended media files. It demonstrates strong capabilities in complex reasoning, mathematics, and code generation. This version of the model is part of a series that includes the smaller Flash and Edge models, with the Core variant serving as the most capable flagship designed for high-stakes and complex tasks.
Reka Core is trained on a mixture of public and proprietary datasets, with support for over 30 languages. It utilizes a modular transformer architecture and was developed with an emphasis on computational efficiency relative to its performance benchmarks in multimodal perception and logical reasoning.