PrismML brings a 27B-parameter AI model to smartphones

PrismML introduces Bonsai 27B, a compressed version of Qwen3.6 27B that the company describes as the first model of its size to run entirely on a smartphone. PrismML announces in an official blog post that the new model shrinks a 27 billion parameter system down to a few gigabytes without abandoning its reasoning, tool use, and vision capabilities.

The release comes in two variants. Ternary Bonsai 27B uses three-value weights and needs 5.9 GB of storage, making it suited for laptops. 1-bit Bonsai 27B uses two-value weights and fits into 3.9 GB, small enough for the memory budget of an iPhone 17 Pro. Both keep the model’s full 262,000-token context window and support multimodal input such as screenshots and documents.

Capability retained, footprint shrunk

According to PrismML’s own 15-benchmark evaluation, Ternary Bonsai 27B retains 95 percent of the original model’s performance, while the 1-bit version keeps 90 percent. Math and coding scores stay close to the uncompressed baseline, while agentic tool calling and instruction following show larger, though still moderate, drops.

The company frames the release as a shift for AI agents that carry out many steps in a row, such as operating software tools or working through documents. Running such workloads locally removes per-step cloud costs and keeps private data on the device, PrismML argues. It also points to hybrid setups combining local models for routine steps with cloud models for harder ones.

Bonsai 27B runs on Apple devices through MLX and on Nvidia GPUs through CUDA. Model weights are freely available under the Apache 2.0 license, and PrismML offers a limited-time developer preview API alongside the release.