Investor Says Free 5.9-Gigabyte Home Model Hits 98 Percent of Its Giant Parent's Performance
It arrived with other powerful open releases from Alibaba and Xiaomi.
All-In with Chamath, Jason, Sacks & Friedberg · Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment FailsA free 5.9-gigabyte AI model you can run on a home desktop delivers "98% of the performance of the big Qwen model" it was built from, investor David Friedberg said on a podcast published Sept. 26.
That model is Bonsai2, released Sept. 17 by Prism ML as a fork of Qwen. Friedberg said it has 27 billion parameters and runs on a Mac Studio or a computer with an Nvidia chip, with no fees beyond running the machine.
Bonsai2 arrived alongside other powerful open releases. Friedberg said Qwen 2.1, released Sept. 20, outperforms Google's Nano Banana 2 image model, and Xiaomi's 309-billion-parameter Mimo Pro, released Sept. 22, is on par with Claude Opus 5 across most benchmarks.
Friedberg claimed that desktop models had gone far beyond the most advanced language model of a year earlier, and in his telling "you don't even need data centers to do 90% of what you can do with AI."
Questions people ask
What does open weights mean?
Open weights means the trained settings, the numbers that make the model work, are published for anyone to download and run. You still need your own computer to run them, but you do not pay per use and you can use them offline.
How can 27 billion parameters fit in 5.9 gigabytes?
It uses compression, a technique called quantization that stores each parameter using fewer bits than full precision. That lets a large model fit in much less memory so it can run on a desktop machine.