Reka Core: Our Frontier Class Multimodal Language Model

We are excited to introduce our largest and most capable model yet, Reka Core.

It is a frontier-class multimodal language model on par with leading models in the industry today. Core was efficiently trained from scratch on thousands of GPUs over a period of a few months.

Performance highlights

Core is competitive with models from OpenAI, Anthropic, and Google across key industry-accepted evaluation metrics. Given its footprint and performance, on a total cost of ownership basis, Core delivers outsized value. The combination of Core’s capabilities and its deployment flexibility unlocks vast new use cases.

Core is comparable to GPT-4V on MMMU, outperforms Claude-3 Opus on our multimodal human evaluation conducted by an independent third party, and surpasses Gemini Ultra on video tasks. On language tasks, Core is competitive with other frontier models on well-established benchmarks.

The table below summarizes a comparison of Core with leading models in the market today.

Source: Reka internal information, company websites, publicly disclosed information, and technical reports. “-” denotes data that is not disclosed or not relevant / applicable due to model capabilities.

Rankings on Human Evaluation for Multimodal.

*A higher ELO score represents better performance. Rankings measured in ELO computed from third-party blind human preferences on a diverse test set of multimodal prompts.*

Capabilities

Multimodal (image and video) understanding. Core is not just a frontier large language model. It has powerful contextualized understanding of images, videos, and audio and is one of only two commercially available comprehensive multimodal solutions.
128K context window. Core is capable of ingesting and precisely and accurately recalling much more information.
Reasoning. Core has superb reasoning abilities (including language and math), making it suitable for complex tasks that require sophisticated analysis.
Coding and agentic workflow. Core is a top-tier code generator. Its coding ability, when combined with other capabilities, can empower agentic workflows.
Multilingual. Core was pretrained on textual data from 32 languages. It is fluent in English as well as several Asian and European languages.
Deployment Flexibility. Core, like our other models, is available via API, on-premises, or on-device to satisfy the deployment constraints of our customers and partners.

While we release a first version today, we expect Core—along with our other models—to continue to break performance barriers as it undergoes further training. Check out our technical report here and example outputs here for more information.

Our partners

In less than a year, Reka has become one of only two developers providing models that allow for comprehensive multimodal input. Its three models allow image, video, and audio input in addition to text. This enables broader and differentiated customer use cases for industries including e-commerce, social media, digital content and video games, healthcare, and robotics, to name a few.

A crucial part of delivering on our mission to make frontier multimodal models that benefit humanity is our various partners. We are proud to count amongst our partners leading global technology platforms and government organizations such as Snowflake, Oracle, and AI Singapore. They enable our customers, organizations, and individuals around the world to benefit from and build with Reka models by democratizing access to multimodal technology.

“We are excited to partner with Reka to bring Reka Core’s impressive industry leading performance to customers through Snowflake Cortex,” said Baris Gultekin, Head of AI, Snowflake. “Snowflake and Reka are at the forefront of AI innovation, and we’re excited to offer the state-of-the-art AI capabilities, all from within Snowflake’s built-in security and governance.”

"Reka’s dynamic multimodal and multilingual models allow enterprises to unlock more value from their data, streamline complex tasks, and realize cost efficiencies,” said Greg Pavlik, senior vice president, AI and Data Management Services, Oracle Cloud Infrastructure. “We look forward to collaborating with Reka to bring video, audio and image capabilities to organizations globally.”

"One of the goals of AI Singapore is to develop highly capable open-source models for Southeast Asia. We are excited about Reka's new models and look forward to working closely with their series of highly performant general-purpose models to build the next generation of SEA-LION models for the region," said Leslie Teo, Senior Director of AI Products, AI Singapore, a national programme launched by the National Research Foundation, Singapore, to anchor deep national capabilities in AI.

Concluding remarks

We live in a multimodal world and Reka is dedicated to the development of frontier multimodal models. The launch of Core today marks an important milestone in delivering against our mission. With our full suite of models comprising Edge, Flash, and now Core, we are ready to take on a greater range of challenges. We are excited by what the future holds and look forward to welcoming new partners who want to join us on that journey