Proton, the privacy-friendly internet service provider based in Europe, has jumped on the LLM bandwagon and released an LLM service last month called Lumo. Besides touting its privacy features, Proton's PR focuses on open source:

Unlike other AI assistants, my code is fully open source, so anyone can verify that it’s private and secure — and that we never use your data to train the model. (source)

Elsewhere on the Proton website, we find a claim that Lumo is "based upon open-source language models" (we are left to guess which). A comparison table shows Lumo alongside some other LLM assistants, with a feature "Opens source code for the public" prominently checked for Lumo and DeepSeek. Factcheck: for DeepSeek that is definitely not the case; at best it is an open weights model and very little is known about its source code. Does Lumo fare any better?

It turns out that Lumo hits a new record in openness: it is the least open "open" AI assistant that we have ever added to our index. One of our reasons for inclusion is an openness claim: a model provider that calls their system "open" or "open source" or a variation on that. Lumo is "open" in that sense (Proton calls it open source) but in no other way. Nothing about it is currently open.

Here is a comparison of the openness status of Lumo and two well-known other models in the space: DeepSeek (included in Proton's own comparison), and OLMo from AllenAI, the current leading openness champion.

In the past, we've gotten model providers like Mistral retreat from using the term "open source" and instead use the more precise "open weights". But Proton will need to do more work to get there, as they don't even reach the middling openness levels of Mistral and other model providers evasive about training data.

It may be that Lumo is a bespoke fine-tuned version of an open weights model like Mistral, OLMo or Llama. It may be a wrapper for an API from Anthropic or another provier. It may be all of these things together in a trenchcoat. The point is: we would know if it were truly open source.

On the sunny side: the only way from here is up. Literally disclosing anything, from training data to model weights to model architecture to data sheets, will be an improvement over the current status. We will be following its rise with great interest.

Update September 1

Proton's Lumo account on BlueSky has responded to this post saying "Lumo isn't a model; it uses open models, including the model listed as the best for openness in this piece. You can find the models Lumo uses in our privacy policy". That page repeats the open source claim:

Lumo’s code is open source, meaning anyone can see it’s secure and does what it claims to. We’re constantly improving Lumo with the latest models that give the best user experience.

The only open source code we have found is for the Lumo mobile and web apps. Proton calling the Lumo AI assistant open source based on that is a bit like Microsoft calling Windows open source just because there's a github repository for Windows Terminal.

The models listed on Lumo's privacy policy page are "Nemo, OpenHands 32B, OLMO 2 32B, and Mistral Small 3". OpenHands is a QWEN fine-tune, and Nemo and Mistral Small are both Mistral models. Since Proton has open-sourced neither the Lumo system prompt nor the mysterious routing methods that decide which model will handle your query, you never know what you are going to get. At best, Lumo can only be as open as the least open system they use. In practice, it has to be even less open than that, because Proton has added additional undisclosed optimization steps and further layers of routing obscurity.

In the absence of technical documentation of system architecture, we cannot update our assessment. If Proton releases more detailed information, we welcome a pull request with the requisite updates. Perhaps the conclusion will be that Lumo is not enough of a unified system to merit inclusion; or perhaps additional information will allow it to rise through the openness ranks. As noted before, none of this would come up if Lumo were actually open source.