It doesn’t say anything about how this works. “No data retention” suggests that data is going to a server that’s processing it but then not retaining it. That’s bad, this type of feature should run entirely on the client.
Also, there’s a proprietary AI model doing the work, even if inference is in the client. It says “open weight” but that’s like letting people download a program as a compiled binary, the exact opposite of what FOSS is supposed to mean. It’s an open/free model only if the training data and algorithms are also public, not just the trained weights.