📰 Key Takeaways

PrismML is a still relatively unknown AI lab that has only closed a $22.25M seed round, but its technical approach and team pedigree are drawing serious attention. The company was founded by a group of Caltech researchers, and CEO Babak Hassibi is a Caltech professor specializing in compression. Its advisor roster includes a Databricks co-founder and Ion Stoica, director of UC Berkeley’s Sky Computing Lab. Investors include Khosla Ventures, Cerberus Capital, and Caltech.

PrismML’s core pitch is that high-performing reasoning LLMs don’t have to be “big.” Last Thursday (September 17), the company released its latest model, Bonsai 2 27B, which compresses Alibaba’s open-source Qwen3.8 27B model down to just 5.9GB — a 9-to-10x reduction in size, small enough to run on a regular PC or even a high-end smartphone (there are rumors PrismML is in talks with Apple, though Hassibi hasn’t confirmed this).

On performance, Bonsai 2 hits 98% of the original Qwen model’s score on combined benchmarks, up from the first-generation Bonsai’s 95% when it launched back in March. That first model has already been downloaded more than 11 million times, with smaller derivative models racking up another 2.6 million downloads.

Technically, PrismML gets there by compressing the model’s “weights” — the information the model learns and stores during training. A typical model needs 16 bits per weight, but PrismML’s “ternary” weight scheme reduces each value down to just three possibilities — +1, −1, or 0 — dramatically cutting storage needs.

On the competitive front, Multiverse Computing, founded by researchers from Spain’s Donostia International Physics Center, is also working on LLM compression and has raised even more funding — see the original article for details.

💬 JudyAI Lab Take

PrismML just surfaced with a $22.25M seed round, a team out of Caltech, CEO Babak Hassibi focused on compression, and Databricks co-founder Ion Stoica on the advisory board.

What’s really turning heads is their technical claim that reasoning models don’t need to be big. The newly released Bonsai 2 27B compresses Qwen3.8 27B down to 5.9GB — a 9-to-10x size reduction — while still holding onto 98% of the original model’s performance, small enough to run on a regular PC or even a phone. The trick: cutting weights from 16 bits down to just three “ternary” values (+1, -1, 0). The first-gen Bonsai has already crossed 11 million downloads, which tells you demand for lightweight models isn’t just hype. This points to a broader trend: as inference cost becomes the bottleneck for AI adoption, compression and quantization techniques are proving just as valuable as scaling up parameter counts. Spain’s Multiverse Computing working on the same problem shows this isn’t a one-off — it’s a direction the whole industry is taking seriously. For AI builders, it’s worth understanding the basics of ternary weights and quantization, and thinking about whether your own models have room to slim down too.

📅 Source Info

🔗 Further Reading