Bonsai 27B: The First 27-Billion-Parameter AI Model That Runs on a Phone

Until recently, running a truly capable AI model on a smartphone was either a fantasy or a compromise. Today, that boundary has shifted: PrismML released Bonsai 27B, the first 27-billion-parameter model that fits entirely on a device you carry in your pocket.

The key is a technique called extreme low-bit quantization. A standard 27-billion-parameter model in 16-bit precision occupies roughly 54 gigabytes — far too large for any phone. PrismML reduced each weight to just 1 bit (binary) or 1.58 bits (ternary), compressing the entire model to about 3.9 GB for the 1-bit version. This is smaller than many modern mobile games.

Remarkably, the model retains most of its intelligence. On a benchmark suite covering math, coding, tool-calling, instruction-following, knowledge, and vision tasks, the 1-bit Bonsai 27B scored 76.1% overall compared to the full-precision baseline of 85.0%. In math and coding — the capabilities most critical for agentic workflows — performance dropped by only 3 to 7 percentage points. The ternary variant, at 5.9 GB, scored 80.5% and runs comfortably on laptops.

The implications extend beyond convenience. When a capable model runs entirely on-device, user data never leaves the machine. Persistent AI assistants can operate offline, reason over private local files by default, and execute multi-step agentic loops at zero marginal cost — no cloud API fees, no network latency, no data leakage. PrismML demonstrated a computer-use agentic loop running on an iPhone 17 Pro Max and a full agentic desktop workflow on an RTX 5090, both using Bonsai 27B.

PrismML introduced the concept of "intelligence density" — capability per gigabyte — to measure this shift. By that metric, the 1-bit Bonsai 27B delivers more than 10 times the density of the full-precision baseline and roughly 2.7 times the best conventional low-bit alternative. As AI workloads move from single API calls to sustained, multi-step agentic work, the economic and architectural case for on-device intelligence grows stronger.