PrismML released Bonsai 2 27B yesterday. It compresses Qwen3.8 27B, a widely used open source model from Alibaba, down to 5.9 gigabytes — nine to ten times less memory than the original needs. Small enough for an ordinary PC, and possibly for a well-equipped phone.
PrismML was founded by Caltech researchers and is run by Babak Hassibi, a professor there and a compression specialist. Ion Stoica advises the company; he co-founded Databricks and directs Berkeley’s Sky Computing Lab. The backers are Khosla Ventures, Cerberus Capital and Caltech itself. There are rumours of talks with Apple, which Hassibi declined to discuss with TechCrunch.
Three values instead of 16 bits
The trick is in the weights. Normally each weight takes 16 bits. PrismML uses ternary weights and gets by with three states: plus one, minus one, zero.
Not a new idea, but apparently a well-executed one. PrismML isn’t alone either — Multiverse Computing in Spain is working on the same problem with considerably more funding.
98 rather than 95 percent
The movement matters more than the absolute figure. Bonsai 2 hits 98 percent of uncompressed Qwen’s aggregate benchmark scores. The first Bonsai, from March, reached 95 percent. The gap to the original shrank from five points to two within one model generation.
The demand is real: the company says the first Bonsai was downloaded over eleven million times, and its even smaller models another 2.6 million. Whether 100 percent is ever reachable, Hassibi leaves open. Compression will probably always cost something.
Those two percent aren’t the problem
Honestly, perfect parity is an academic question. No model is that accurate uncompressed, and no benchmark reflects real work closely enough for two percentage points to show up in daily use. What matters far more is the software around it — the harness the model runs inside.
It still matters to us, for a reason that has nothing to do with benchmarks. A usable reasoning model running locally on your own machine answers a question people ask about every agent they run: where does the data go? The answer ‘nowhere’ is worth more than any benchmark point. I’ll trade two percent of performance for that any day.
Sources: