AI
PrismML's Compression Breakthrough Shrinks Advanced AI Models Down to Smartphones
A Caltech-founded startup has found a way to squeeze high-performance language models into just gigabytes of space, potentially bringing sophisticated AI to everyday devices.
TechCrunch — AI
Caltech researchers are tackling one of AI's fundamental challenges: how to make powerful reasoning models lean enough to run directly on consumer hardware. PrismML, a startup led by compression expert Babak Hassibi, is pursuing a radical approach — one that could reshape how people access artificial intelligence.
The Compression Feat#
The company recently unveiled Bonsai 2 27B, a model that condenses Alibaba's Qwen 3.8 27B down to just 5.9 GB — a roughly tenfold reduction in file size. This means the model can now fit on standard laptops and potentially high-end smartphones, a dramatic shift from the sprawling infrastructure typically required to run such systems.
The magic lies in a technique called ternary weights. Rather than storing each weight (the learned parameters that make up a model) as a standard 16-bit number, PrismML reduces them to just three possible values: +1, −1, or 0. This radical simplification slashes memory demands without substantially degrading performance.
Minimal Performance Loss#
What sets PrismML apart from other compression efforts is its ability to preserve capability. The newest Bonsai variant maintains 98% of the original model's benchmark scores — a significant jump from the first version released in March, which retained 95%. The initial Bonsai model has already notched over 11 million downloads, signaling real-world adoption of the approach.
According to Hassibi, "virtually no performance" is lost in the process. The modest gaps that remain may be largely academic, since uncompressed models themselves are imperfect and benchmarks don't always reflect how systems perform on actual tasks.
Looking Ahead#
PrismML plans to extend this technique to far larger models in the coming months — potentially hundreds of billions of parameters. Hassibi suggests that bigger models may be even better candidates for compression, since their sheer scale creates more "redundancy" to eliminate.
The startup, backed by Khosla Ventures, Cerberus Capital, and Caltech, counts advisor Ion Stoica (a Databricks co-founder and director of UC Berkeley's Sky Computing Lab) among its team. Though unconfirmed, the company has been reported to be discussing partnerships with Apple.
