Bonsai 27B Just Became the First 27B AI Model to Run Directly on Your Phone
Quick Read:
- PrismML has released Bonsai 27B, the first 27-billion-parameter AI model built to run natively on a smartphone.
- It comes in two versions: Ternary Bonsai 27B (5.9GB, laptop-optimized) and 1-bit Bonsai 27B (3.9GB, phone-optimized).
- The model is based on Qwen3.6 27B and retains roughly 90–95% of the original model’s performance despite the massive compression.
- Everything is open-sourced under the Apache 2.0 license, with support for Apple devices via MLX and NVIDIA GPUs via CUDA.
Local AI just crossed a threshold that seemed technically out of reach until now. PrismML has officially announced Bonsai 27B, and the headline claim is a big one: it’s the first 27B-class model that can actually run on a phone. Not a stripped-down, “lite” version, but a genuine 27-billion-parameter model doing real reasoning, tool use, and multimodal work, sitting in someone’s pocket.
To understand why this matters, it helps to know what usually stands in the way. A 27B model at standard 16-bit precision needs around 54GB of memory. Even an aggressively quantized 4-bit version still needs about 18GB, which rules out phones and most laptops entirely. That’s been the wall for local AI: you could shrink an 8B model down small enough for a device, but anything in the 27B range, with all the extra reasoning power that scale provides, stayed locked to the cloud or high-RAM machines.
Bonsai 27B, built on top of Qwen3.6 27B, is PrismML’s answer to that wall. The release ships in two variants designed for two different jobs. Ternary Bonsai 27B uses a three-value weight system and lands at 5.9GB, aimed at laptop-class quality. The 1-bit Bonsai 27B variant goes further, compressing down to 3.9GB with 1.125 effective bits per weight, specifically built to fit inside the memory a phone app is actually allowed to use.
That Last Point Is Worth Sitting With
A phone doesn’t hand its full memory over to a single app. A 12GB iPhone might only offer around 6GB to a model, and that budget also has to cover the KV cache and other overhead. Most conventional low-bit builds of a 27B model don’t come close to fitting inside that. Bonsai 27B does, with room left over.
What’s notable is that this compression doesn’t gut the model’s actual usefulness. Across a 15-benchmark suite covering math, coding, tool calling, and vision, Ternary Bonsai 27B retains around 95% of the full-precision baseline, while the 1-bit version holds onto roughly 90%. Math and coding scores barely move. Tool-calling and agentic benchmarks take a somewhat bigger hit, but for a model running on a phone, that’s still a meaningful jump forward from anything previously possible at this scale.
That Agentic Angle Is Really The Point Of Bonsai 27B
PrismML isn’t positioning this as a lightweight chatbot for simple questions. It’s built for multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops, the kind of tasks that require a model to plan, execute, and correct itself across many steps. Because it’s multimodal, Bonsai 27B can also process screenshots, documents, and camera images directly on-device, supporting up to roughly 262,000 tokens of context.
Running a model like this locally changes the cost and privacy equation in a real way. When the model lives on the device, a hundred-step agentic loop costs nothing extra to run, and the user’s data never has to leave the phone or laptop. That opens the door to persistent on-device assistants, offline agents, and tools that can safely reason over private data because there’s no round trip to a server at all.
PrismML has released the full Bonsai 27B model weights today under the Apache 2.0 license, alongside a free, limited-time developer preview API. It runs natively on Apple devices through MLX and on NVIDIA GPUs through CUDA, using custom low-bit kernels built around its hybrid-attention architecture. For a company that emerged from a team of Caltech researchers with backing from Khosla Ventures, Cerberus, Google, and Samsung, Bonsai 27B looks like the clearest proof yet that the intelligence-per-gigabyte curve is bending sharply in favor of local, on-device AI.
