They don’t, but with quantization and distillation, as well as fancy use of fast ssd storage (they published a paper on this exact topic last year), you can get a really decent model to work on device. People are already doing this with things like OpenHermes and Mistral (given, 7B models, but I could easily see Apple doubling ram and optimizing models with the research paper I mentioned above, and getting 40B models running entirely locally). If the start of the network is good, a 40B model could take care of a vast majority of user Siri queries without ever reaching out to the server.
For what it’s worth, according to their wwdc note, they’re basically trying to do this.
They don’t, but with quantization and distillation, as well as fancy use of fast ssd storage (they published a paper on this exact topic last year), you can get a really decent model to work on device. People are already doing this with things like OpenHermes and Mistral (given, 7B models, but I could easily see Apple doubling ram and optimizing models with the research paper I mentioned above, and getting 40B models running entirely locally). If the start of the network is good, a 40B model could take care of a vast majority of user Siri queries without ever reaching out to the server.
For what it’s worth, according to their wwdc note, they’re basically trying to do this.