Opportunity assessment: high. A Caltech spinout just proved you can shrink a frontier-class open model roughly 10x and keep 98% of its brains.
On Thursday, PrismML released Bonsai 2 27B, a compressed version of Alibaba’s Qwen3.8 27B that fits in 5.9 GB, according to TechCrunch AI. That’s small enough for a PC and, possibly, a high-end smartphone. The startup has raised only a $22.25 million seed round, but TechCrunch AI reports it’s rumored to be in talks with Apple. CEO Babak Hassibi declined to comment on that.
Situation report
- The product. Bonsai 2 27B is the second model in PrismML’s Bonsai family. It cuts memory 9x to 10x versus the original Qwen model while matching 98% of its aggregate benchmark scores.
- The trajectory. The first Bonsai shipped in March at 95% parity. Two months later, 98%. That’s a real improvement curve, not a one-off demo.
- The traction. The original Bonsai has been downloaded over 11 million times. PrismML’s even smaller models add another 2.6 million downloads, per the company.
- The people. Hassibi is a Caltech professor and compression expert. Ion Stoica, Databricks co-founder and director of Berkeley’s Sky Computing Lab, is an adviser. Backers include Khosla Ventures, Cerberus Capital, and Caltech.
How the compression works
A model’s weights are the information it learns during training. Normally each weight takes 16 bits of storage. PrismML uses “ternary” weights instead: every weight becomes +1, -1, or 0. Three possible values per weight instead of 65,536. Store far less per weight, and the whole model shrinks dramatically.
The hard part isn’t the idea. Ternary weights have been floating around research circles for years. The hard part is doing it without lobotomizing the model. That’s PrismML’s claim to fame: Hassibi says its LLMs have “lost virtually no performance” compared with the originals.
Why 98% is good enough
I’ll be honest, the 2% gap is mostly academic. Uncompressed LLMs aren’t perfectly accurate in the first place. Benchmarks don’t perfectly reflect real tasks either. And the harness a model runs inside matters a lot for real-world accuracy. A 2% benchmark drop is unlikely to change how the model performs on your actual work.
Hassibi also thinks the gap closes as models get bigger. “The next models that we will release, hopefully in the next couple of months, will be in the several-hundred-billion-parameter range, and I expect it will be easier to retain the intelligence there,” he told TechCrunch. His logic: larger models have “more room to be able to compress them without losing the intelligence.”
The competitive picture
PrismML isn’t alone here. Multiverse Computing, founded by a professor from Spain’s Donostia International Physics Center, works on the same problem and has raised far more money. The difference PrismML is selling is fidelity. Anyone can shrink a model. Shrinking it while keeping 98% of the performance is the claim that matters.
Why this matters
What stands out here is the economics. Every query you send to a cloud model costs somebody money and leaks your data to a third party. A 27B reasoning model running locally flips both of those.
Stoica put it plainly:
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.
That’s a direct threat to the API-per-token business model, and a direct gift to anyone building products that need to run offline, on-device, or under strict privacy rules.
What to do with this
- If you build on open models, test Bonsai 2 against your current Qwen deployment. A 10x memory cut can turn a GPU-only workload into a laptop workload.
- If you ship consumer apps, watch the Apple rumor. On-device reasoning at this size is the missing piece for private assistants.
- If you sell inference, start pricing in a world where a chunk of the workload never leaves the user’s device.
The next test is whether the several-hundred-billion-parameter release lands in the promised couple of months and holds the 98%+ line. If it does, small stops being a compromise and becomes the default. Full details are in the TechCrunch AI report.