In an innovative expansion of its technology offering, the artificial intelligence startup Reka has just launched Reka Core, an advanced multimodal language model capable of processing texts, images, audio, and video. This new development not only matches but in some cases surpasses other leading models in the market in various benchmark tests.
Reka Core, developed from scratch in less than a year, has been trained using licensed public and synthetic data. This model stands out for its 128,000-token context window and its multilingual capabilities in 32 languages, setting a new standard in the industry.
In performance evaluations, Reka Core has demonstrated superiority over Claude 3 Opus in multimodal human evaluations, has outperformed Gemini Ultra in video tasks, and has shown competitive performance with GPT-4 in image comprehension.
The benchmark tests of Reka Core are impressive, and competing head-to-head with the best models in the market from scratch is no small feat. With comprehensive multimodal capabilities and a short training period, Reka is positioned as a major new player in the field.
Reka Core is part of Reka’s Firefly model family, which includes images, vectors, layouts, and text effects. Additionally, Reka is working on integrating third-party AI models from companies like OpenAI, Runway, and Pika Labs into its platform.
The model not only handles video tasks but also understands and processes images and audio, positioning it as one of the two commercial multimodal solutions currently available.
Reka has formed strategic partnerships with companies such as Oracle and Snowflake, in addition to collaborating with governmental organizations like AI Singapore. These collaborations are designed to democratize access to multimodal technologies and expand the reach and applicability of Reka’s models.
With a memory capacity of 128 GB and a memory bandwidth of 3.7 TB, Core is specially equipped to handle large language and multimodal models, increasing the performance of workloads and data center cost efficiency.
Core’s encoding capabilities and agent workflows position it as a top-tier code generator, capable of empowering complex multilingual workflows, essential for tasks requiring sophisticated analysis.
The introduction of Reka Core represents a significant milestone in the company’s mission to offer cutting-edge models that facilitate the creation and production of content in an increasingly multimodal world. With this launch, Reka not only aims to tackle broader challenges but also invite new partners to join them on their journey towards continuous innovation in the field of artificial intelligence.