PrismML, a Caltech spinout with a $22.25 million seed round behind it, released Bonsai 2 27B on Thursday. The model is a compressed version of Alibaba's open Qwen3.8 27B that fits in 5.9 GB, nine to ten times less memory than the original needs, and scores 98% of the original's aggregate benchmark results. The number to watch is not 5.9 GB. It is 98 — up from 95% for the first Bonsai, which the company shipped in March. A three-point gain in roughly six months is the actual claim here, and it is a claim about a trend line rather than a product.
That distinction matters because compression only becomes interesting at the point where the loss stops being a reason to say no. At 95%, a compressed model is a compromise you accept when you have no alternative. At 98%, the argument flips: the remaining gap is small enough that the surrounding software — the runtime, the scaffolding, the prompt handling — plausibly moves accuracy more than the compression does. CEO Babak Hassibi, a Caltech professor whose specialty is compression, does not claim the gap ever closes. He expects some loss to remain permanent.
The technique is ternary weights. A model's weights are what it absorbs and stores during training, and storing each one normally takes 16 bits. PrismML forces every weight to be one of three values: +1, −1, or 0. A fuller description sits on the project's GitHub page. Do the arithmetic on the numbers PrismML published and Bonsai 2 lands at roughly 1.75 bits per parameter — close to the floor for a three-valued scheme, which leaves very little room for a Bonsai 3 to be smaller rather than merely more accurate.
The company was founded by a group of Caltech researchers and is advised by Ion Stoica, co-founder of Databricks and director of the Sky Computing Lab at UC Berkeley, the lab that produced Letta and SGLang among others. Khosla Ventures, Cerberus Capital and Caltech itself are the backers. PrismML is not alone in the category: Multiverse Computing, founded by a professor at the Donostia International Physics Center in Spain, works on the same problem and has already raised significant funding.
Distribution is the strongest card PrismML holds. The first Bonsai has been downloaded more than 11 million times, with another 2.6 million downloads across the company's smaller models. For a startup whose entire seed round is $22.25 million — modest by the standards of what AI infrastructure companies have been raising — that is an unusual ratio of reach to capital.
It is also the part of the story that needs the most scepticism. Downloads are not revenue, and nothing in the announcement explains how free weights on a model hub turn into a business. The plausible answer is licensing to whoever ships the hardware, which is why the unconfirmed talks with Apple are the most commercially loaded line in the whole story. Hassibi declined to comment on them to TechCrunch, and a declined comment is not a confirmation. But it does point at the shape of the company: PrismML looks less like a model lab than like a compression vendor that needs one large device maker to say yes.
There is a structural exposure underneath that. PrismML does not train models. Bonsai 2 is a derivative of someone else's open release, which means the product line depends on Alibaba continuing to publish Qwen openly — and on Alibaba not shipping competent quantized weights itself. The defensible asset is the method, not the artifact, and methods described on a public GitHub page are a thinner moat than a 98% benchmark score suggests.
Notably absent from the announcement: any figure for speed. Everything PrismML published is about memory. Fitting a 27-billion-parameter model into 5.9 GB says it can be loaded on a laptop or a phone; it says nothing about how fast it runs once loaded, and "runs on device" and "usable on device" are not the same statement. The 98% is also an aggregate across benchmarks, which conceals whether the reasoning tasks — the specific capability PrismML sells — degrade more than the average. A composite that holds while its most important component slips is exactly the shape a composite is good at hiding.
The next step is the one that tests everything. Hassibi says PrismML's following models could arrive within months at several hundred billion parameters, and that larger models should be easier to compress without losing intelligence — more redundancy to remove, more room to reach 100% of the original. If he is right, the company has already done the hard version first and the roadmap gets easier from here. But take Bonsai 2's ratio and apply it to a 300-billion-parameter model and you get something in the range of 65 GB. That is my arithmetic, not the company's, and it is the tension in Ion Stoica's pitch that on-device models mean no subscription and no data leaving the device: the models PrismML is most confident it can compress cleanly are the ones least likely to fit on the device that was supposed to run them.