PrismML shrinks Qwen3.8 27B model to 5.9GB by storing each weight as +1, -1 or zero

The file is 5.9GB ^[1]. That gets a reasoning model into a laptop's storage. What decides whether a feature ships is what the operating system hands your process at the moment a user taps the button, with a browser and a video call already resident. TechCrunch's report describes the model as small enough for a PC and, possibly, a high-end smartphone ^[5]. The hedge on the phone case is the report's own.

Take the 27B label as 27 billion weights. At the usual 16 bits each that is 54GB, so 5.9GB is a 9.2x cut, consistent with the 9x to 10x PrismML claims ^[20][4]. Divide the other way and 5.9GB is about 47 gigabits, or roughly 1.75 bits per weight ^[21]. Three states need log2(3), about 1.58 bits, as a floor, which puts the shipped file around 0.17 bits a weight above the minimum for pure ternary storage ^[22].

Bonsai 2 keeps 98 percent of Qwen3.8 27B's aggregate benchmark score, three points up from the first Bonsai's 95 ^[6][7][23]. The comparison is to the open-source model it compresses, not to whatever frontier API a team pays for now. TechCrunch argues the remaining two points are academic, since uncompressed models are not that accurate, benchmarks are not that reflective of real tasks, and the harness a model runs inside affects accuracy as well ^[17]. Hassibi says compression will likely always have some impact ^[18].

"There is more room to be able to compress them without losing the intelligence," Hassibi told TechCrunch about larger models. "So I would just say, as a general trend, for larger models, it's easier to get to 100%" ^[11].

PrismML says the first Bonsai has been downloaded more than 11 million times and its smaller models another 2.6 million, 13.6 million in total ^[8][19]. Teams read counts like that as demand for on-device inference. A download shows that a developer pulled a file.

Ion Stoica, a PrismML advisor and director of Berkeley's Sky Computing Lab, said: "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's also going to be private, because you're not going to send it to the cloud" ^[12][14]. Free covers the compute bill. TechCrunch's report does not include tokens per second, latency or battery draw for the model on any device ^[24].

The company has raised a $22.25 million seed from Khosla Ventures, Cerberus Capital and Caltech ^[13]. Hassibi, a Caltech professor who works on compression, declined to comment to TechCrunch on rumored talks with Apple ^[15]. Multiverse Computing, founded by a professor from Spain's Donostia International Physics Center, is working on the same compression problem with considerably more money raised ^[16].

Does the request need frontier quality, or is a 27B-class answer good enough. Does the prompt have to stay on the device. Local wins one quadrant: a 27B-class answer is good enough and the data cannot leave, and there Bonsai 2 replaces a network call. Where a 27B-class answer is good enough and the cloud is fine, the choice is between hosting Qwen yourself and pushing 5.9GB down every user's connection on every version bump ^[1]. Where the request needs frontier quality, the token bill stays where it is.