Vinod Khosla sits in the background of Apple’s latest AI search. Oliver reported that Apple has already held meetings with PrismML about using its model-shrinking technology to bring more advanced AI to the iPhone.
The number that makes the report matter is stark: PrismML says it compressed the 54GB Qwen 3.6 model to 4GB. That kind of shrinkage points to a much larger model fitting on a phone instead of living on a server.
PrismML’s 54GB to 4GB cut
PrismML is a startup that uses mathematical wizardry to shrink the size of large language models, and it says the Qwen 3.6 model has 27 billion parameters. Phone models normally have just a few billion parameters, so the reported jump is not cosmetic.
PrismML also says the model it made work on an iPhone can handle complex chat and reasoning. It can also power fully autonomous agents, which is the clearest sign that Apple is looking past simple text generation and toward more capable on-device AI.
Apple and PrismML meetings
Apple wants bigger AI models on the iPhone, and it prefers advanced models to run on-device. The company has been limited by the huge amount of data more advanced models require, which is why a smaller footprint is the practical barrier here.
Oliver wrote that Apple has identified PrismML as a potential partner in its AI-shrinking efforts. He also wrote, “I would guess Apple will just buy the company if it's the real-deal.”
Google, Khosla Ventures and Cerberus Ventures
Google provided PrismML with computing resources and monetary grants used to train the 1-bit Bonsai model family on Google V4 TPUs, and Google is apparently continuing that support. Investments in PrismML are coming from Khosla Ventures and Cerberus Ventures, which gives the startup more than one route to keep building while Apple decides whether the talks become something larger.
Whether Apple and PrismML turn those meetings into a formal partnership is still the open question, and the report says iPhone and Siri AI users do not yet have a timeline for any benefit.
