The Caltech spinout claims it shrunk a 54 GB model to under 4 GB, and Apple is reportedly paying attention
Running a 27 billion-parameter AI model on a smartphone sounds like claiming you fit a grand piano into a backpack. PrismML, a Caltech spinout backed by Khosla Ventures, says it did exactly that, compressing Alibaba’s Qwen 3.6 large language model from roughly 54 GB down to less than 4 GB and running it on an iPhone 17 Pro with all parameters active.
The company emerged from stealth on March 31, 2026, armed with $16.25 million in seed funding led by Khosla Ventures and a technology that could reshape how AI gets deployed on consumer devices.
What PrismML actually built #
PrismML’s proprietary compression techniques, showcased through its Bonsai model family, achieve up to 14x smaller memory footprints and 8x faster inference compared to full-precision models. The model gets radically smaller and runs radically faster, while keeping every single parameter active rather than selectively pruning the model’s capabilities.
CEO Babak Hassibi has emphasized that this technology could enhance AI capabilities without requiring the kind of data center investments that have turned AI infrastructure into a multi-hundred-billion-dollar arms race.
Why Apple cares, and why crypto should too #
Apple has reportedly been in discussions with PrismML about integrating its compression technology into future AI offerings. Apple’s current on-device model, the AFM 3 Core Advanced, tops out at 20 billion parameters with a sparse architecture. PrismML’s technology could potentially let Apple leapfrog that ceiling without redesigning its chip architecture.
Running a full-capability LLM on your phone, with no data leaving the device, aligns with crypto’s foundational argument that users should control their own data. It also potentially reduces the demand for the cloud compute tokens that several crypto projects have built their entire business models around.
PrismML’s Bonsai models demonstrate significant energy savings compared to running equivalent models on server infrastructure.
What this means for investors #
The $16.25 million seed round led by Khosla Ventures is modest by AI startup standards, where seed rounds routinely stretch into nine figures. The claims themselves remain largely unverified by independent benchmarks.
For crypto-native investors specifically, decentralized compute networks like Akash, Render, and io.net have attracted capital based on the premise that AI inference demands will overwhelm centralized providers. On-device inference at this scale doesn’t eliminate that demand entirely, but it could meaningfully compress the addressable market for cloud-based AI compute, particularly for consumer-facing applications. What’s worth watching is whether Apple formalizes its relationship with PrismML and what that means for the next generation of iPhone AI features.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our