PrismML, a startup founded by researchers from Caltech and backed by Khosla Ventures, Cerberus and Google, has announced Bonsai 27B, the first artificial intelligence model with 27 billion parameters (27B) capable of running directly on a mobile phone. And not only that: it does so with just 3.9 GB in size, a technical feat that until now seemed impossible.
Two versions for two scenarios
Bonsai 27B is based on Qwen3.6 27B and comes in two flavors:
- Ternary Bonsai 27B (5.9 GB): uses ternary weights {−1, 0, +1} with FP16 group scaling, offering 1.71 effective bits per weight. It is the quality-oriented variant, ideal for laptops, with multi-turn reasoning, tool calling and agent loops.
- 1-bit Bonsai 27B (3.9 GB): uses binary weights {−1, +1}, with 1.125 effective bits per weight. It fits within the memory budget of an iPhone 17 Pro, bringing a 27B-class model to a phone for the first time.
Both variants are multimodal (they process text, images, documents and camera) and support a 262K token context with speculative decoding to speed up generation.
And how well does it work?
The benchmarks speak for themselves: the ternary version retains 95% of the original model’s performance in full precision (FP16), and the 1-bit version retains 90%. In critical areas such as math and code, the loss is minimal. In agentic reasoning and tool calling, it stays within a few points.In perspective: Bonsai 27B in 1-bit reaches up to 163 tokens/second on an NVIDIA RTX 5090 and up to 87 tokens/s on an Apple M5 Max. On the iPhone 17 Pro Max, the model runs completely on-device.
The key: intelligence density
PrismML has spent years refining neural network compression techniques without sacrificing reasoning capability. Its founder explains it with its own metric: “intelligence density” (intelligence per GB). With 1-bit Bonsai 27B, you get 0.53 per GB: more than 10 times the density of the original model in FP16, and about 2.7 times the best available low-precision alternative.The model is distributed under the Apache 2.0 license and is already available on Hugging Face. It runs natively on Apple (Mac, iPhone, iPad) via MLX and on NVIDIA GPUs via CUDA. PrismML also offers a free developer API for a limited time.
Why it matters
Until now, a 27B model in 16 bits took up about 54 GB, and in 4 bits about 18 GB —too much for a phone. A 12 GB iPhone only exposes about 6 GB usable for the model, and that’s while sharing with the KV cache and activations. No conventional build of a 27B model even came close. The 1-bit Bonsai 27B, at 3.9 GB, is the first to get through that door with plenty of room to spare.High-end artificial intelligence is no longer tied to the cloud and expensive servers. It fits in your pocket. And that changes the rules of the game for AI privacy, latency and accessibility.
