Reka Core, the company’s first multimodal language model, was released by AI startup Reka AI. According to the company’s statement, Core is “competitive” with top models from OpenAI, Anthropic, and Google when it comes to industry-accepted metrics and can work with images, audio, and video.
Reka AI is a San Francisco-based startup with two Southeast Asian founders, and it was started by researchers from DeepMind, Google, and Meta. The chief scientist of the company is Yi Tay, a Singaporean, and the CEO, Dani Yogatama, is an Indonesian.
Reka Core is a large language model with “powerful” contextualized image, video, and audio understanding. According to the company, there are only two developers whose commercially available models support comprehensive multimodal input, the other being Google.
Reasoning skills, such as language and math, can be used for tasks requiring complex reasoning thanks to Core’s 128,000-token context window. Moreover, it is a code generator that enhances agent-based workflows.
Reka AI asserted that Core performs better on video tasks than Google’s Gemini Ultra, another multimodal model, when compared to comparable models available in the market. According to some, Core is also similar to GPT-4V on MMMU, an AI benchmark that assesses a model’s ability to complete tasks requiring reasoning and knowledge appropriate for a university.
Reka AI released its Edge and Flash models a few months prior to the release of Core. Together with Radical Ventures and DST Global Partners, the company has raised US$58 million in funding as of June 2023. The last known valuation of Reka AI is $300 million, according to market research platform CB Insights.
Reka AI’s partners include Snowflake, Oracle, and AI Singapore. On the basis of valuation per employee, it was recently ranked fifth among all AI firms worldwide.