'Bonsai 27B,' an AI model with 27 billion parameters that can be run locally on an iPhone, has been released. It is a smaller version of Qwen3.6-27B with a memory usage of 3.9GB, allowing it to run at a practical speed on a smartphone alone.

AI development company PrismML has officially announced 'Bonsai 27B,' a compressed version of its 27 billion parameter AI model that can run on smartphones. PrismML explains that even the 1-bit version, with a file size reduced to 3.9GB, maintains approximately 90% of the performance of the original 'Qwen3.6-27B.'

PrismML — Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone

https://prismml.com/news/bonsai-27b

Running AI on a device makes it less susceptible to network conditions and allows for the processing of personal and business data without transferring it from the device. However, high-performance AI models require a large amount of memory. Saving Qwen3.6-27B with typical 16-bit precision results in a file size of approximately 54GB, and even a version with reduced precision to 4 bits still requires about 18GB. This is a size that is difficult to handle not only on smartphones but also on many laptops.

PrismML claimed to have compressed Qwen3.6-27B to less than 4GB in early July 2026 and made it run on an iPhone 17 Pro. As of July 10, 2026, the compressed version was scheduled to be released on July 14, and this Bonsai 27B is the official release of the model that was previously announced. Details such as the model name, performance evaluation, and distribution method have now been revealed.

The 27 billion parameter AI 'Qwen3.6-27B' runs on the iPhone 17 Pro; Apple is reportedly in discussions with PrismML and compression technology - GIGAZINE

Bonsai 27B comes in two versions: '1-bit Bonsai 27B' which prioritizes memory efficiency, and 'Ternary Bonsai 27B' which prioritizes quality. 1-bit Bonsai 27B represents weights as '-1' and '+1,' with an effective precision of 1.125 bits per weight, including correction data. The model capacity is 3.9GB, which is said to fit within the memory capacity available to apps on the iPhone 17 Pro. Ternary Bonsai 27B uses a three-value system: '-1,' '0,' and '+1,' with an effective precision of 1.71 bits and a theoretical capacity of 5.9GB.

In an evaluation that combined 15 benchmarks including knowledge, reasoning, mathematics, programming, instruction response, tool calling, and image understanding, the original Qwen3.6-27B scored 85.0 points, while Ternary Bonsai 27B scored 80.5 points and 1-bit Bonsai 27B scored 76.1 points. PrismML states that the performance retention rates compared to the original models were approximately 95% and 90%, respectively.

Beyond simply reducing storage capacity, Bonsai 27B is designed for applications such as processing long documents and performing multiple tasks sequentially. It can handle up to approximately 262,000 tokens of context and accepts not only text but also screenshots, documents, and camera images as input. It also supports features for calling external tools in a defined format and processing for AI agents that can proceed to the next operation while reviewing the results of their work.

The operating speed is reported to be up to 163 tokens per second for the 1-bit version and up to 134 tokens per second for the 3-level version on the NVIDIA GeForce RTX 5090. On the Apple M5 Max, it was reported to be up to 87 tokens per second for the 1-bit version and up to 58 tokens per second for the 3-level version. Apple products use MLX, while NVIDIA GPUs use CUDA.

The new PrismML's Bonsai 27B model running on iPhone 17 Pro

Bonsai 27B is the first 27B-class model to run on a phone and it uses less than 5GB of memory

Pushing the limits on what can run a phone pic.twitter.com/utAoEE3zON

— Adrien Grondin (@adrgrondin) July 14, 2026

As of the time of writing, the Bonsai 27B model data is released under the Apache License 2.0, and versions such as the MLX version for Apple products and the GGUF version usable with llama.cpp are distributed on Hugging Face.

Bonsai 27B - a prism-ml Collection
https://huggingface.co/collections/prism-ml/bonsai-27b

The company stated that PrismML's compression techniques are not limited to specific AI structures, and that they are working on applying them to even larger models and new structures. They concluded their announcement by saying, 'Early computers filled rooms, but now they fit in your pocket. AI is following the same path, and Bonsai 27B is the biggest step forward yet.'