Reka AI launches multimodal language model to rival Google’s Gemini

Reka Core ranks second in the ELO rating system / Photo credit: Reka AI

AI company Reka AI has launched Reka Core, its first multimodal language model. Working with images, audio, and video, Core is “competitive” with leading models from OpenAI, Anthropic, and Google across industry-accepted metrics, the company said in a statement.

Reka AI – which was founded by researchers from DeepMind, Google, and Meta – is based in San Francisco but has two Southeast Asian founders. Singaporean Yi Tay is the company’s chief scientist, while CEO Dani Yogatama grew up in Indonesia.

In addition to being a large language model, Reka Core has “powerful” contextualized understanding of images, videos, and audio. The company said that it is one of just two developers – the other being Google – whose commercially available models allow for comprehensive multimodal input.

Core has a 128,000-token context window, which enables reasoning abilities – including language and math – suitable for tasks that require complex reasoning. It is also a code generator that can empower agentic workflows.

In comparison to similar models in the market, Reka AI claimed that Core outperforms Google’s Gemini Ultra – also a multimodal model – on video tasks. Core is also said to be comparable to GPT-4V on MMMU, an AI benchmark that evaluates how models perform tasks that require university-level knowledge and reasoning.

Core’s launch comes a few months after Reka AI rolled out its Flash and Edge models. The company has raised US$58 million so far, which came in June 2023 from a round led by DST Global Partners and Radical Ventures. Market intelligence platform CB Insights pegs Reka AI’s last known valuation at US$300 million.

Snowflake, Oracle, and AI Singapore are among Reka AI’s partners. It recently ranked fifth among global AI firms on a valuation per employee basis.

See also: This Singapore aquafarm takes to AI like fish to water