AI research lab PrismML released a language model Thursday that compresses a widely used AI system down to just 5.9 GB—small enough to run on personal computers and potentially high-end smartphones. The new model, called Bonsai 2 27B, represents a ninefold to tenfold reduction in memory compared with the original version, according to a report published by TechCrunch on September 17. The startup, founded by Caltech researchers and led by professor Babak Hassibi, is betting that powerful reasoning models don't actually need to be massive to perform well.

The latest release compresses Qwen3.8 27B, an open source model from Alibaba, while retaining 98% of the original's aggregate benchmark scores. That's an improvement over PrismML's first Bonsai model, launched in March, which matched 95% of performance. The original has been downloaded more than 11 million times, with the company's even smaller models racked up another 2.6 million downloads, PrismML says. The firm raised $22.25 million in seed funding and counts Khosla Ventures, Cerberus Capital, and Caltech among its backers. Ion Stoica, co-founder of Databricks and director of Berkeley's Sky Computing Lab, serves as an adviser.

PrismML achieves the compression by shrinking the "weights" that make up a model—essentially the information a system learns and stores during training. Normally, each weight requires 16 bits. The startup's approach, called "ternary" weights, simplifies that to three values: +1, −1, or 0. Hassibi told TechCrunch that perfect benchmark parity is "fairly academic anyway," since language models aren't perfectly accurate in their uncompressed form and benchmarks don't perfectly reflect actual tasks. A 2% degradation likely wouldn't meaningfully affect how a model performs in real use, he said. The company is also rumored to be in talks with Apple, though Hassibi declined to comment on that.

The startup's next goal is to apply its compression technique to even larger models in the several-hundred-billion-parameter range, which Hassibi expects to release within the next couple of months. As model size grows, there's more room to compress them without losing intelligence, making it easier to reach 100% performance parity with larger systems, he told TechCrunch. Stoica explained that the technology enables advanced models to run on users' devices, meaning "you are going to have intelligence at your fingertips, and it's going to be free because it's going to run on the device you already bought." It will also be private, since data won't need to be sent to the cloud. For enterprises weighing device-based inference against cloud deployments, the performance trade-off narrows considerably as compression techniques mature. Companies betting exclusively on centralized infrastructure may find themselves at a strategic disadvantage if on-device capabilities continue closing the quality gap.