Proton launches a privacy-friendly AI called Lumo
Proton has launched its own AI chatbot. Say hello to Lumo, the privacy-friendly AI.
Proton's announcement talks about how AI services are turning people into products, but also that AI can be useful when it is private.
Proton says that Lumo does not store any logs of your conversations on its server. The chats that you save are store on your device and encrypted. Speaking of which, the encryption implemented in Lumo is the same open-source tech that is used in Proton Mail, Drive, etc. The Switzerland-based company promises that it will not share any user data to third-parties, and that Proton and Lumo will not user your conversations to train the large language model. Its servers are based in Europe.
Proton also took a swipe at Apple, when it highlighted that it has not partnered with OpenAI or other AI companies based in America or China. In case you aren't aware, Proton filed a lawsuit against Apple last month regarding the App Store's policies, calling them unfair and anti-competitive.
Now, coming to Lumo, it uses 4 Large Language Models: Nemo, Mistral Small 3, OpenHands 32B, OLMO 2 32B (Allen Institute for AI).
You don't need an account to access the chatbot, but you will need one to save your conversations. Lumo's Ghost mode prevents chats from being saved. Lumo can search the web to fetch answers related to your query, offer writing assistance, and can even analyze files that you upload, and those stored in your Proton Drive. Currently, it only supports documents, so you can't use it with images or videos. On that note, I should point out that Lumo seems to be a text-exclusive chatbot, i.e. it cannot generate images or videos.
During my tests, I asked it questions about various topics, and it said that it does not have real-time browsing capabilities or have access to databases to fetch information from. That's a bit of a letdown, I suppose this explains why the AI doesn't need to be trained, it is designed for specific purposes like those I mentioned above.
You can try Lumo for free at https://proton.me/lumo-home or download the mobile app for Android and iOS.
EDIT: The above URL no longer works. Use https://lumo.proton.me/guest instead
(Thanks Boris and Tom for the update.)
The free version has some limitations, but users can optionally get Lumo Plus which offers unlimited daily chats, multiple large uploads, extended chat history and priority access to the bot.
Here is a chart that compares Lumo's privacy practices versus those of Google Gemini, Microsoft Copilot, OpenAI ChatGPT and DeepSeek.
Normally, comparisons that portray a company's service as superior to its rivals is considered unethical, but given that the others are absolutely terrible for privacy, I think it is perfectly ethical.
No logs, no training, proper encryption, this is what AI services should have done right from the start. What do you think?
“Normally, comparisons that portray a company’s service as superior to its rivals is considered unethical”
What ?
Not really a criticism, I suppose it’s useful for casual users interested in keeping their questions private.
But as an alternative, one can get the same level of quality and privacy by running a local AI model, provided you have a decent-sized GPU or a fairly high-powered CPU with plenty of RAM.
I use an AMD 5950X with 64GB of RAM and a crappy old AMD GPU with only 2GB of VRAM ( the GPU useless for AI use.) I can run models up to around 12-14B parameters and while they are somewhat slow, they work. They can pretty much answer any questions the Proton models can – if you get the “good” local models, like the Google Gemma, Microsoft Phi and some others.
All you need is a GUI front end like AnythingLLM, LMStudio, or my favorite, MSTY. Some of them will let you chat with multiple models at once (if you have enough RAM), and compare answers.
You can also get some local vision models, although quality is fairly crappy.
I know you can run AI on cpu. BUt 5950 is a top end cpu for AM4.I have the lessor 5800x and something like Llama runs farely well but Deepseek struggles quite a bit. I only know about the cpu performance because the localllm software fails to acknowledge my GPU and it’s a known bug by the developers that’s supposed to get fixed. Also AMD GPUs are starting to get better AI support.
Regrettably, it is also neutered AI. My first test for AI is to ask some provocative or ridiculous questions and see if it dares to respond. This one failed my test. Only safe questions/answers. And that’s too bad. For AI that is supposed to be privacy safe, it would have been a big benefit if AI could answer spicy/unsafe questions.
If you do not want an account, go directly to [https://lumo.proton.me/guest].
[https://proton.me/lumo-home] does not redirect and gives “Page not found”.
@boris,
> “[https://proton.me/lumo-home] does not redirect and gives “Page not found”.]
Now, indeed, not when I had opened the link sooner today. A recent change obviously.
Good you point it out. [[https://lumo.proton.me/] it is, only, with ‘guest’ if no account.
Nice as well reading your experience with informal queries.
@Tom Hawack
And I did not even ask anything that illegal in my country. Just ridiculous and unsafe (Politically incorrect, sex jokes, curse words and so on). This is my test how many guardrails AI has. And it’s not like I am looking for AI girlfriends. I know that I can ask few serious questions a day to ChatGPT/Grok and that enough for me. But it would be nice to have free private AI that I can ask something that I can not ask somebody I know.
Proton will do anything except updating their core products with much needed features. I don’t need standard notes, proton wallet, or another private AI (Have they not heard of Venice & Leo?). They’re trying to do too much and their original products are all suffering for it.
Interesting for me as well.
Note :) “Lumo’s Ghost mode prevents chats from being saved.” : rather “Guest” mode I guess. Funny Freudian slip (not sure it’s relevant according to Freudian analysis though). In French we have a word for that : “lapsus” which corresponds more to “slip” and we’d evoke “lapsus révélateur” (relevant slip) to implicitly tie it to any “Freudian” mechanism… even if we’d rather refer to Carl Jung in this case …
Back to Proton’s
BTW, [https://proton.me/lumo-home] redirects (here) to [https://lumo.proton.me/guest] (“guest” of course given no account).
I just asked Lumo if it accepted search queries via a dedicated url parameter: the answer was “no”. Of the very feww AI Chats I very seldom open (Deepfind AI Search, DuckDuckGo AI Chat, Mistral Le Chat) Deepfind is the only which accepts a search query, but be noted it presents itself and indeed is an “AI Search” engine” and not an “AI Chat” bot.
Nice to have Proton’s AI but as far as I’m concerned not sure what it offers without an account will increase the near to zero number of connections I establish with above mentioned competitors. I’ll have four instead of three. But, frankly, not my cup of tea, which is definitely not an argument : guess I prefer to search than to find (when Picasso stated in an interview that needed not to search given he’d simply find : maybe the characteristic of a genius?).
DGG’s AI is optional, and includes your query when you activate it. I don’t like the other AI’s.
I am going to check out Proton’s AI. It would should still be good for cutting down on sorting through a long list of hits, helping with online shopping, getting travel deals and options, and similar things.
@NotOnMySideOfTheCloud, true that DDG’ AI includes your query when you activate it… given you activate it from DDG. Personally I deactivate DDG AI on DDG Search (as I do, e.g. with Brave Search’s AI) given I separate Search and AI as a principle (an arguable principle). Yet, what I may appreciate is calling DDG AI from anywhere with a search query. But I’m aware that an AI Chat is not an AI Search, queries become questions and therefor are generally more consistent than a word or two. At the end this wish becomes secondary, sort of invalid I guess.
I have checked Proton’s AI and even with the reported advantages of having an account, I haven’t been thrilled.
I think that for an extra-occasional user as myself, no AI audio/video in view, DDG AI is quite suitable. Has brought me already interesting assitance of which finding the right REGEX when I’d face too complex schemes for my technically uneducated brains… yet, funny thing is that the logic of REGEX sort of makes it way in our reasoning, in our ability to determine them without us being aware of it … really odd!
“No logs, no training, proper encryption, this is what AI services should have done right from the start. What do you think?”
What does it take for a blog writer to be hired by a site like “TechSpot” or “PCMag”?
No “resume” (logs of personal history); “no education,” especially in grammar, punctuation, etc.; “distrust” of authority figures and a lack of the proper understanding of information that can and cannot be shared.
If employers modeled requirements for employment based on the requirements set forth for AI models proposed in the article, what would happen?
Would destroy the application process–employers would have to “guess” at every applicant and determine at an interview–no they couldn’t have one of those because of “no logs” and “encryption.” Hmmm . . . would be a mess and completely illogical.
Interesting. Like that you have a guest mode and are not forced to sign in to an account.