Multimodal large language model research startup Reka AI Inc. today announced it has raised $110 million in fresh funding backed by Nvidia Corp. and Snowflake Inc.

Founded in 2022, Reka develops multimodal models focusing on ultra-efficient training and serving infrastructure for delivering inference at low cost.

The company offers multiple AI models including Reka Flash, a multimodal model capable of understanding video, images, text and audio, which is the flagship of the company’s family of models. Reka Vision and Reka Research are the company’s agentic AI platforms that have recently entered general availability providing complex question answering and visual search capabilities.

According to a report from Bloomberg, the current funding round triples the company’s valuation to $1 billion, up from about $300 million after the company raised $50 million in a funding round that included the participation of Snowflake’s venture capital arm.

Last year, Snowflake made overtures to acquire Reka, reportedly at a price more than three times higher than its valuation at the time. According to Bloomberg, the deal talks were scrapped after both companies decided to move on independently, but they still hold ties today.

Vivek Raghunathan, vice president of AI engineering at Snowflake, told Bloomberg that Reka would provide AI models and other services to the company.

“Very few teams in the world have the capability to build what they’ve built,” Raghunathan said. “Almost everyone at that level of talent is at OpenAI, Meta or Anthropic. Reka is one of the rare independents — and they’ve proven they can compete.”

Reka recently released a competitively performing version of Reka Flash 3.1 that the company said shows high performance against similarly sized models such as Quen3-32B and OpenAI’s o3-mini.

Reka Flash itself is a compact 21 billion-parameter model, which are the internal variables used by models to make decisions and also shape the overall size of a model impacting where it can fit. Using a method called quantization, a model’s memory footprint can be reduced, allowing it to run with less resources, such as mobile or edge devices.

According to the company, Reka Quant allows for almost lossless quantization when processing Reka Flash 3.1, bringing down the size significantly with very little performance loss. According to the company Reka, Quant reduces resource usage with only a 1.6-point average degradation on hard reasoning benchmarks. The next best quantization routine incurs a 6.7-point average degradation.

The company has grown from 20 to 50 employees over the past year. Although initially focused on building efficient models, Reka also develops tools to assist companies with AI interfaces and application logic.

Reka said it intends to use the investment to significantly accelerate technical development on its models and scale its multimodal platforms, aiming for wider enterprise adoption and expansion of its global reach.

Image: Pixabay