This startup builds robot brains with open-source grit

A war is raging between the proponents of open-source and closed-source AI. Amid this landscape, one Singapore startup is making a splash.

Menlo Research wants to build the “brain for robots,” software that helps machines see, think, and act in the real world.

Image credit: Timmy Loen

The company is behind Jan, a lightweight desktop app that lets users run large language models locally, without sending any data to the cloud. 

Since launching in October 2023, Jan has exploded in popularity. The free app has been downloaded millions of times, with an average of 7,000 new installs per day. Some users have touted it as an easy way to run open-source LLMs without the need for powerful machines. 

Menlo also goes all-in on open source, making its solutions publicly available. And it’s doing it all without venture funding even as investors continue hunting for more AI golden eggs.

But what exactly is Menlo Research and can a bootstrapped outfit really keep open-source ambitions alive?

“Server is busy” no more

The company was co-founded by husband and wife Daniel Ong and Nicole Zhu, both Stanford alums. Ong was previously the CTO of student loan platform Danacita, while Zhu served as vice president of fraud and risk at Gojek.

Ong insists Menlo, which has a 27-member team, isn’t founder-centric. “The real investors are the people who chose to work here,” he tells Tech in Asia, noting that several team members left high-paying roles – including at companies like Tesla – to join Menlo.

To back that up, Menlo distributes 1% of company equity every six months to team members, including part-time staff and open-source contributors. Equity is based on performance, with recent cycles covering as many as 18 people.

Much of Menlo’s popularity today comes from Jan. The app can load popular open-source models – like Alibaba’s Qwen3 or DeepSeek R1 – in just a few clicks and run them entirely offline.

Once installed, Jan allows users to download popular open-source models such as Qwen3 or DeepSeek-R1 / Image credit: Menlo Research

Based on Tech in Asia’s tests, a standard laptop can handle the app, even without a discrete graphics processing unit, though performance improves significantly with one. On a mid-range gaming PC, for instance, DeepSeek R1 running inside Jan can process each prompt in under a minute.

“Don’t break too many robots” 

The open-source dilemma