The Illusion of "Full Power"
Getting a model this large onto consumer hardware means you have to make some serious compromises. PrismML's approach is aggressive quantization. They've got two main variants:
- Ternary Bonsai 27B: This one uses weights that are {−1, 0, +1}. It's effectively 1.71 bits per weight, and the model size drops to 5.9 GB. This is their "quality-oriented" version, aimed at laptops.
- 1-bit Bonsai 27B: This is the real headline grabber. Binary weights {−1, +1}, an effective 1.125 bits per weight, and a tiny 3.9 GB model size. This is the one that fits the ~6 GB app memory budget of an iPhone 17 Pro.
They claim 95% benchmark retention for the Ternary variant and 90% for the 1-bit. Those numbers sound solid on paper. But when you look at the actual scores against the Qwen 3.627B baseline, you see where the intelligence starts to bleed out.
Where the Bits Fall Apart
Here's what matters about those benchmark numbers. The full Qwen 3.627B baseline scores an 85.0 overall. The Ternary Bonsai drops to 80.5. The 1-bit Bonsai? It's down to 76.1.
That's a 9-point drop from the baseline for the "phone-ready" version. And it's not evenly distributed.
The Cool Part vs The Dealbreaker
| Capability | Ternary Bonsai (vs. Baseline) | 1-bit Bonsai (vs. Baseline) | My Take
