Apple has identified PrismML as a potential partner in its effort to shrink AI models for the iPhone, and the two have already held meetings about how the startup’s technology could bring more advanced AI to Apple devices. The interest matters because Apple wants larger models to run on-device, not in the cloud, and PrismML says it has already compressed a 54GB model to 4GB without hurting performance.
That makes Vinod Khosla a name traders and AI watchers are searching for now, because Khosla Ventures is among the investors backing the company. The report places Apple’s interest in PrismML at a moment when the iPhone maker is trying to move beyond smaller, less capable models and into systems that can handle more ambitious tasks directly on the phone.
PrismML’s example is the Qwen 3.6 model, which has 27 billion parameters and was reduced from 54GB to 4GB. The company says the compressed version still works on an iPhone and can handle complex chat and reasoning, while also powering fully autonomous agents. In practical terms, that means the math is doing the work: by lowering the storage footprint without changing the output behavior, the model becomes light enough to fit within the limits of a device that has far less room than a server. Apple has been limited by the huge amount of data that more advanced models require, which is why the promise of running larger systems on the phone is drawing attention.
The friction is that Apple has wanted better on-device AI for some time and has already run into performance problems with its own model-shrinking efforts. PrismML is betting that its approach avoids that tradeoff entirely, and that is what makes the company interesting to Apple. Google has also supplied PrismML with computing resources and monetary grants used to train the 1-bit Bonsai model family on Google V4 TPUs, and that support is apparently continuing, while investments are also coming from Khosla Ventures and Cerberus Ventures.
What Apple does next is still unclear. The meetings show it is looking seriously at outside help, but it is not yet public whether the relationship becomes a partnership, a purchase, or just another test of whether PrismML can do on iPhones what Apple has been trying to do on its own.
