Bonsai 27B: New AI model works directly on iPhone 17 Pro Max without cloud

The PrismML startup presented the Bonsai 27B multimodal model, which is able to fully work on the Apple iPhone 17 Pro Max smartphone without the need to connect to cloud servers. This is an important step in the field of local data processing using AI.

Bonsai 27B is a specialized model based on the Qwen3.6 27B architecture. It supports a wide range of tasks, including working with text, image recognition, programming, invoking tools and executing multi-step instructions.

Technical characteristics and weight of the model

PrismML engineers have significantly optimized the weight of the model so that it can function in the RAM of a mobile device.

• Binary version (1 bit): the volume is only 3.9 GB.
• Ternary version (1.58 bit): provides better quality, using 5.9 GB of memory.

For comparison, a standard 16-bit model of the same level would require about 54 GB of disk space and RAM.

Performance across devices

At the moment, the company provides the following measurements of generation speed:

• iPhone 17 Pro Max: about 11 tokens per second.
• Mac with M5 Max chip: up to 87 tokens per second.
• NVIDIA GeForce RTX 5090 video card: up to 163 tokens per second.

It is important to consider that the basic figures of 3.9 GB refer only to the “language” part. When working with visual content, an additional 0.6–0.9 GB is required for a special module. In total, running the model taking into account the context of 4000 tokens consumes about 5.9 GB of memory on the iPhone.

Accuracy of work

According to PrismML’s own tests, the optimized versions show high performance compared to the Qwen3.6 27B reference model:

• Binary version retains up to 90% accuracy.
• The ternary version demonstrates up to 95% efficiency (according to 15 key benchmarks).

Versions and availability

The model is not a ready-made program for the average user. It works on Apple devices thanks to the modified MLX environment. The weight and deployment tools are already open source on the Hugging Face platform under the Apache 2.0 license.