A 27 Billion Parameter Reasoning Model Now Fits in 5.9 Gigabytes

By Omar aka Crypt0 · Sep 17, 2026

While the industry pours concrete for gigawatt data centers, a team of Caltech researchers just proved the most exciting AI frontier fits in your pocket. PrismML released Bonsai 2 27B on Thursday, a reasoning model with 27.8 billion parameters compressed down to 5.9 gigabytes. That is more than nine times smaller than the full precision model it descends from, and it runs on a laptop, a desktop, and, with room to spare, a modern smartphone.

The compression trick is called ternary weights. Every weight in the model, the numbers it learned during training, normally takes 16 bits to store. PrismML shrinks each one to three possible values: plus one, minus one, or zero. Across the whole model that works out to 1.76 effective bits per weight, applied end to end. The result keeps 98.2 percent of the aggregate benchmark performance of Qwen3.8 27B, the open Alibaba model it is built on, up from 95 percent for the first Bonsai release back in March.

The deployment story is what matters. The model runs natively on Apple hardware through MLX and on NVIDIA GPUs through CUDA, ships with a 262,000 token context window, takes text and image input, and is released under the Apache 2.0 license with a free developer preview API. The original Bonsai 27B has already been downloaded more than 11 million times, with another 2.6 million downloads for the smaller models, and Bonsai 2 lifts capability retention from 95 percent to 98.2 percent.

The people behind it explain the pedigree. CEO Babak Hassibi is a Caltech professor and a compression specialist, and adviser Ion Stoica co founded Databricks and directs Berkeley's Sky Computing Lab. The startup raised a $22.25 million seed round from Khosla Ventures, Cerberus Capital, and Caltech. Hassibi told CNBC that Apple is evaluating the technology, measuring its speed, energy efficiency, and on device performance. The lab calls its thesis intelligence density: progress measured in capability per gigabyte, per watt, per dollar.

Stoica puts the stakes plainly. Intelligence at your fingertips, free because it runs on the device you already bought, and private because your queries stay on your own hardware instead of traveling to a distant server farm. That is the self ownership story in a single sentence. A model this capable running locally means your assistant, your coder, your researcher, all working for you with your data staying yours.

The roadmap points larger, which sounds paradoxical until Hassibi explains it. The next releases, expected within the next couple of months, will compress models in the several hundred billion parameter range, and he expects retention to climb even closer to 100 percent there, since bigger models carry more redundancy to squeeze. If the trend holds, the gap between frontier capability and pocket sized deployment keeps shrinking.

What this unlocks for you is simple: the most personal computer you own, your phone, is about to become the most private AI lab you own too. Download a model, run it offline, ask it anything, and the cloud stays out of the loop entirely. The data center builders are racing to rent you intelligence by the hour. PrismML is betting you would rather own it outright.

Sources

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.