PrismML has released Bonsai 2 27B, a compressed language model that fits 5.9 GB onto a PC or high-end smartphone.
The startup does not rely on venture capital for its current trajectory. It has secured only a $22.25 million seed round. Its focus remains on the technical minds behind the project and the potential of the technology itself.
Small models, big claims
PrismML argues that capable, reasoning models do not require massive size. The company is building systems small enough to run on personal computers and mobile phones. CEO Babak Hassibi declined to comment on reports that Apple is in talks with the firm.
The latest release compresses Qwen3.8 27B, an open source model from Alibaba, down to 5.9 GB. This represents a nine to ten times reduction in memory usage compared to the original.
Hassibi leads the company as a professor at Caltech and an expert in compression technologies. The founding team includes Caltech researchers. Ion Stoica serves as an adviser. He co-founded Databricks and directs Berkeley’s Sky Computing Lab, which has launched several technologies and startups, including Letta and SGLang.
Investors backing the startup include Khosla Ventures, Cerberus Capital, and Caltech.
Other companies, such as Multiverse Computing, are also working on LLM compression. Multiverse Computing was founded by a professor from Spain’s Donostia International Physics Center and has raised significant funding.
Hassibi states that PrismML’s compression technology is unique because the resulting models lose virtually no performance. Bonsai 2 matches 98% of Qwen’s aggregate benchmark scores. The first Bonsai model, released in March, matched 95%. That original model has been downloaded over 11 million times. PrismML’s smaller models have seen another 2.6 million downloads, the company says.
This indicates an improvement in compression results from one release to the next. Whether the technology can ever reach 100% benchmark performance parity remains to be seen. Hassibi notes that compression will likely always have some impact.
Perfect benchmark parity is largely academic. LLMs are not perfectly accurate in their uncompressed form, and benchmarks do not perfectly reflect actual tasks. A 2% degradation would likely not meaningfully affect how a model performs in real use. The surrounding software, which determines how a model runs, also matters significantly for accuracy.
PrismML achieves this by shrinking the weights that make up a model. These weights are the information a model learns and stores during training. Normally, each weight requires 16 bits. PrismML’s approach, called ternary weights, simplifies this to three values: +1, −1, or 0. With smaller values to store for each weight, the model takes up dramatically less space. For details on the compression technique, see the project’s Hugging Face page.
Future goals
The startup’s next goal is to apply this compression technique to even bigger models. Hassibi told TechCrunch that the next models, hopefully released in the next couple of months, will be in the several-hundred-billion-parameter range. He expects it will be easier to retain intelligence there.
As model size grows, there is more room to compress them without losing intelligence. Hassibi added that as a general trend, it is easier to reach 100% parity with larger models.
Stoica says he is excited for this technology because it makes advanced models runnable on users’ devices. Intelligence will be at your fingertips and free because it runs on the device you already bought. It will also be private because you will not send it to the cloud.
