PrismML shrinks a 27B AI model to 3.9 GB—your iPhone just became a data center 📱

PrismML has released Bonsai 27B, a 27-billion-parameter AI model small enough to run on a consumer smartphone, with a binary variant weighing 3.9 GB and a ternary variant at 5.9 GB. The binary build runs at 11 tokens per second on an iPhone 17 Pro Max, while the ternary version reaches roughly 26 tokens per second on an M5 Pro laptop. Both are released free under the Apache 2.0 license. A standard 27B model at half precision typically requires around 54 GB of memory, putting it out of reach for most laptops and many desktops.

The compression method, built on intellectual property from Caltech, reduces each model weight from 16 bits of floating-point precision to a single sign—either +1 or -1 in the binary build, or one of three values in the ternary. Each group of 128 weights shares a 16-bit scaling factor, placing the binary variant at 1.125 bits per weight, roughly 14 times smaller than the full-precision original, while the ternary version settles at 1.71 bits. Unlike conventional quantized models that keep sensitive layers such as embeddings, attention, and the language model head at higher precision, Bonsai compresses those components end-to-end with no escape hatch.

Bonsai 27B is the second major release in the family, following the March launch of Bonsai 8B, a 1.15 GB model that demonstrated the 1-bit architecture could survive at 8 billion parameters without collapsing in reasoning. The jump to 27 billion targets the scale at which sustained chain-of-thought reasoning, reliable tool use, and multi-step agentic behavior emerge consistently.

Across 15 benchmarks evaluated in thinking mode on NVIDIA H100 GPUs—spanning knowledge, math, coding, and tool use—Ternary Bonsai 27B averaged 80.49, or 94.6% of the full-precision model, while the 1-bit variant reached 76.11. On AIME25 and AIME26, modeled on the American Invitational Mathematics Examination, Ternary Bonsai 27B scored 93.7%, with the binary version trailing in the same evaluation bracket.