PrismML, a startup founded by Caltech researchers and led by Babak Hassibi, is drawing attention not because of massive fundraising (it has a $22.25 million seed round) but for the technical team and the compression technology it is developing. The company is betting that high-performing, reasoning-capable large language models do not necessarily need to remain large.
PrismML aims to make reasoning LLMs small enough to run on personal computers and, possibly, high-end smartphones. On Thursday the company released Bonsai 2 27B, a new member of its Bonsai family that compresses Alibaba’s widely used open-source Qwen3.8 27B down to 5.9 GB — roughly a 9x–10x reduction in memory compared with the original. Hassibi declined to comment on media rumors that PrismML is in talks with Apple.
Founders, advisers and backers
PrismML was founded by a group of Caltech researchers and is led by Babak Hassibi, a Caltech professor and an expert in compression technologies. Ion Stoica, co-founder of Databricks and director of Berkeley’s Sky Computing Lab, is an adviser. Investors in the startup include Khosla Ventures, Cerberus Capital and Caltech.
Compression while retaining performance
PrismML says Bonsai 2 retains about 98% of Qwen3.8’s aggregate benchmark scores. That is an improvement over the first Bonsai release in March, which matched about 95%. According to the company, the original (larger) Bonsai model has been downloaded more than 11 million times, and PrismML’s even smaller models have accounted for another 2.6 million downloads.
Hassibi notes that compression will likely always have some impact on performance, and whether perfect parity with an uncompressed model is achievable remains an open question. He and the company argue, however, that a small percentage difference on benchmarks is unlikely to meaningfully degrade real-world usability in most cases; the software environment that runs a model also matters for accuracy.
Ternary-weight technique
PrismML’s approach reduces the storage needed for the model’s weights — the information a model stores during training. Typically, each weight uses 16 bits. PrismML applies a ternary-weight scheme that restricts each weight to one of three values: +1, −1 or 0. Storing these far smaller discrete values dramatically reduces a model’s file size. The project’s Hugging Face page provides more technical detail.
Plans to scale to much larger models
The company’s next objective is to apply its compression method to much larger models. Hassibi told reporters that in the next couple of months PrismML hopes to release models in the several-hundred-billion-parameter range, and he expects it will be easier to preserve intelligence at that scale because larger models offer more room for compression without losing capabilities.
Competition and implications
PrismML is not alone in pursuing LLM compression: other teams, such as Multiverse Computing, are also working in this space and in some cases have raised significantly more capital. PrismML emphasizes that its approach yields minimal performance loss.
Ion Stoica highlighted the practical implications: enabling advanced models to run locally on user devices means intelligence will be available without recurring cloud cost and with improved privacy, since user data need not be sent to remote servers.
Conclusion
Bonsai 2 27B’s reduction of Qwen3.8 27B to 5.9 GB while retaining roughly 98% of aggregate benchmark performance marks a meaningful step toward making powerful LLMs runnable on-device. PrismML plans to test the approach on much larger models in coming months, and the outcome will influence how widely local, private model inference can spread.
