Proton released Lumo — its privacy-focused chatbot built on open-source models — in mid-July, and, following an update in August that addressed some early issues, I find myself using it more often than ChatGPT or Claude. In a world where internet companies have done so much damage to our society, I'm trying to find more ethical tools. But when the competition offers flashier features in exchange for the low price of user data, does Proton seriously think it can compete?
If Eamonn Maguire, Proton's head of machine learning, shares my concerns, he doesn't show it during the hour or so we chat over Zoom. If anything, I start to see why he believes Lumo has found a valuable niche.
Proton began work on Lumo last year following the release of Scribe. The email writing tool was the company's first foray into AI. According to Maguire, the reception to Scribe was "better than [Proton] thought it would be." Soon enough, people were asking the company why the tool wasn't included in their Unlimited subscription. Internally, Scribe also changed how Proton thought about AI. Proton released its mail client in 2014, almost exactly 10 years after the debut of Gmail in 2004. Proton Drive, meanwhile, arrived in 2020 or eight years after Google Drive. The company felt it couldn't be late to chatbots too.
Advertisement Advertisement
"We knew we needed to move faster on this because it's going to be a big privacy problem in the future," said Maguire, pointing to the monetization arc of past platforms like Gmail. After cornering the email market, Google turned to ads and selling user data to fund Gmail's operations. It's a familiar scenario, and one Maguire argues we're already seeing play out again with AI chatbots, with Elon Musk recently telling advertisers that xAI would display sponsored responses alongside regular content from Grok. In other words, the enshitification of AI chatbots has already begun.
The growing strength of open-source models is one of the reasons Maguire believes Proton can compete against the likes of OpenAI and Google. Open systems, particularly those coming out of China, may lag behind proprietary models in user adoption but they're beginning to match them in testing benchmarks. For instance, Zhipu AI's GLM-4.5 currently sits top 10 overall on LMArena. At the same time, "all the top models are starting to cluster together in benchmarks now," Maguire says. Even accounting for some of the safety issues with Chinese models "Overall, open systems are competing, not just tailgating [closed models]."
Lumo employs a combination of smaller open-source models that require fewer resources to run — specifically Nemo, OpenHands 32B, OLMO 2 32B, and Mistral Small 3. Maguire argues Proton's approach allows it to be nimble. That's important for a company that hasn't raised venture capital funding, and needs to think about building a sustainable business model from the start. "A lot of people think they need the best model to get the best responses. But I think it's been shown that you can get very capable responses from models that are smaller," Maguire said.
For consumers, that also means Proton can offer Lumo for less. Like most chatbots, basic features are available for free, with the option to remove rate and token limits through a paid subscription. For Lumo, that costs $13 — less than $20 per month most AI companies charge for their basic paid plans, and substantially less than the $200 tier many of them have begun offering for nearly unlimited usage of their models.
Advertisement Advertisement
Maguire suggests thinking about it this way: You could drive a Formula One car to the grocery store, but that would be overkill. If you find LLMs useful in your workflow, there's a good chance a state-of-the-art model like OpenAI's o3 is similarly overkill for your needs. Unless you're a researcher, you probably don't need a system that can reason over many minutes to tackle a complex problem.
... continue reading