California-based AI startup PrismML announced on July 14 the release of Bonsai 27B, which it says is the first model of its parameter class to run natively on a smartphone.
Built on Alibaba’s Qwen3.6 27B, Bonsai 27B comes in two compressed variants, according to the company’s announcement. A “ternary” version uses three-value weights and needs 5.9GB of storage, retaining 95% of the full-precision model’s performance across 15 benchmarks. A more aggressive 1-bit version shrinks the model to 3.9GB — small enough to fit on an iPhone with room to spare — while keeping about 90% of baseline performance.
Why compression matters
A standard 27B-parameter model typically requires around 54GB of memory, putting it well out of reach of phones and most laptops. PrismML’s low-bit quantization technique cuts that footprint by roughly 90%, the company said, without the steep drop-off in reasoning quality that heavy compression usually causes.
In category breakdowns PrismML published, the ternary version scored 93.4 on math tasks and 86.0 on coding, against a full-precision baseline of 85.0 overall. The models support a 262,000-token context window and run through custom low-bit kernels — natively on Apple devices via MLX, and on Nvidia GPUs via CUDA. On an Nvidia RTX 5090, the 1-bit version reaches up to 163 tokens per second; on Apple’s M5 Max chip, up to 87 tokens per second.
PrismML, founded by a team with Caltech roots and backed by Khosla Ventures, Google and Samsung, released the weights under the permissive Apache 2.0 license, letting developers download, modify and redeploy them commercially without restriction.
The release lands as AI labs race to push capable models onto phones and laptops rather than relying solely on cloud inference — a shift driven partly by data-center capacity constraints and partly by demand for offline, private AI assistants.