Snowflake Inc. is reportedly holding talks to acquire Reka AI Inc., a venture-backed provider of large language models.

Sources told Bloomberg late Thursday that a deal could value the startup at more than $1 billion. The news comes a few weeks after Reka debuted a new LLM, Reka Core, that it says is competitive with frontier models from OpenAI and Anthropic PBC. The company says it trained the algorithm over several months using thousands of graphics cards. 

Reka was founded in 2022 by former AI researchers from Alphabet Inc. and Meta Platforms Inc. Last year, the LLM developer raised a $50 million funding round that included the participation of Snowflake’s venture capital arm. According to Reuters, the investment valued Reka at about $300 million.

That Snowflake is now reportedly holding acquisition talks at a price more than three times higher suggests significant optimism about the LLM developer’s technology. The prospective acquisition price is particularly notable given that Reka likely has limited sales. As of last June, the month its $50 million funding round was announced, the LLM developer wasn’t yet generating revenue. 

This week’s report about the acquisition talks comes two months after Reka debuted its newest AI model. Reka Core is a multimodal LLM that can process not only text but also images, videos and audio. It’s capable of ingesting prompts with up to 128,000 tokens, units of data that can each hold a few letters or numbers.

Reka’s engineers compared the LLM against GPT-4 and Anthropic’s top-end Claude 3 Opus model in a series of internal tests. One of the evaluations used MMLU, a benchmark datasets that comprises several thousand questions spanning dozens of topics. Reka Core answered 83.2% of the questions correctly, which put it about 3% behind GPT-4 and Claude 3 Opus.

According to the company, the LLM managed to outperform Clause 3 Opus in a different benchmark test that compared the models’ multimodal data processing capabilities. Additionally, it proved more adept than Google’s Gemini Ultra at certain video analysis tasks.

Reka offers Reka Core alongside two less capable but more hardware-efficient models. The first is called Reka Edge and features a lightweight architecture that allows it to run on devices with limited onboard computing capacity. A third model, Reka Flash, is situated between the company’s two other LLMs on the price-performance scale.

Snowflake would gain more than just LLMs in the event the acquisition talks lead to a deal. After announcing its funding round last year, Reka disclosed that it has developed proprietary model distillation technology. 

Model distillation is a technique that allows researchers to transfer parts of an advanced LLM’s knowledge to a smaller, more hardware-efficient model. The latter AI subsequently gains the ability to perform some of the tasks that its more advanced counterpart supports. Using a smaller LLM to perform processing tasks lowers hardware requirements and thereby cuts costs.

MosaicML Inc., another venture-backed provider of LLMs and model development technology, was acquired last year by Snowflake rival Databricks Inc. in a $1.3 billion deal. Databricks used the software it obtained through the transaction to launch an AI training toolkit called Mosaic AI Training. The acquisition also helped Databricks developed its recently introduced DBRX open-source LLM.

Image: Unsplash