Lumo 2.0 Max is actually a viable AI to use now, even without signing in, you can attach multiple pdfs and ask a fair number of questions. I would say this is out-performing both Ask Brave and duck.ai from my testing, especially given that it can handle attachments, very cool.
I know that it’s not actually e2ee, but of all the providers, including the ones mentioned above, I trust Proton the most to actually delete what I send through once it’s been processed.
Who here is using it? If you have been using it for a while, have you noticed it get a lot better recently?
Yes, I’m using it even though I’ve stopped paying. Proton needs to fix its subscription payment model once and for all; it’s broken. I want to be able to pay for my Proton Duo subscription annually and Lumo Plus monthly.
The app is also quite slow, like other Proton products. But it works well and, as you say, is better than other products in terms of output quality. It uses GLM 5.3, one of the best open weights models out there.
I’ve been going back and forth between Proton’s Lumo 2.0 Max and DDG’s GPT-5.6 Luna, trying to see which one works better for my needs. Personally, I find that the latter adheres better to my initial formatting prompts (such as “do it this way, in this specific format”) and is a little less rambling and more straight to the point. I do mostly use it for grammar and punctuation corrections for massive texts.
I also find duck.ai more comfortable for light scripting assistance and technical support for certain applications. But, all in all, I think they’re pretty comparable.
I’ve used it a lot recently to build a new server and deploy around 10 services on said server. It had one wobble where a mistyped CLI command deleted all my docker stacks , but since then it’s been great.
Really long chats can lead to context length errors so if that becomes likely I ask for a state summary and start a new chat. This works pretty flawlessly. Every now and then a garbled response comes out but it’s obvious when that happens and a single reply is usually all it takes to get it back on track.
Overall it’s been extremely capable and allowed me to do something I would have otherwise struggled with. I don’t feel like I’ve missed out by not using a frontier model and I love the privacy benefits that come with it.
It’s like a night and day difference compared to what it once was with 1.0, but unfortunately, it still hasn’t been able to replace what I need for work. I do like the improvements they’ve made, but for some reason I still get occasional hallucinations. It’s also still pretty lazy, and I have to poke at it multiple times before it finally gives me what I’m looking for. I think the lazy aspect part of it is still the worst thing about it.
And for some reason, even though I have custom instructions, it’s still extremely argumentative with me. Even when it tells me something that I know for sure is wrong, it wants to argue with me about it until I keep pushing, it finally stops being lazy, actually looks into it, and discovers the correct answer.
On one hand, one of my biggest problems with AI has always been that they’re too much of a yes man. With Lumo though, it’s like I’m talking to a Redditor some times.
Yes, I use it daily since v2.0, last spring, and I am very satisfied, whereas I used to hate it before.
To address some concerns mentioned in previous answers here above, I wrote in the general instructions that it should go straight to the point instead of contextualising every thing, be humble and quickly question its own answers when I push back, and to remove final “safety” recommendations and final “artificial” questions inviting me to enter follow-up prompts.
Why do you say it’s not actually e2ee? I am interested in knowing more. I -maybe naively- assumed it was totally e2ee as all their other products.
LOL Are you saying Redditors are argumentative? At least Lumo won’t ban you for doing nothing wrong and you don’t have to beg it to let you post something like on the joke that is r/privacy. Reddit, the heart, I mean fart, of the internet.
It really is bizarre, and I don’t understand what causes it to behave that way. Maybe because it was trained on Reddit data? But yeah, I can tell Lumo something is factually wrong with what it told me in the most polite way, “Hey, are you sure about that? I think this is wrong, and here’s why” and it’ll still dig its heels in.
Then I’ll explain it again, point out exactly where it’s mistaken, and basically force it to fact check itself. Even after several rounds of that, it’ll sometimes keep arguing before it finally goes, “Oh, yeah, you’re right.”
Like I said, my problem with AI’s before is they were trying to appease you too much and were yes men and would just say you’re correct even if you were wrong. This is the other direction.
That’s the part that drives me insane. And I’m sure every time I speak to it now, it tries to downvote me.
I’ve had the same experience as @Menkork. It does argue more instead of being yes-man and to me this is actually refreshing. Frustrating at times, but at least I know it won’t always validate everything I say.
It’s still unnecessarily argumentative when it is flat-out wrong, and answers in Chinese frequently, but what drove me away completely were the numerous times it said that it has “ignored project information and direct behavior instructions” when responding. Paid models should do as instructed.
, but since then it’s been great.