Prism ML has launched Bonsai 27B, a new multimodal LLM based on Qwen3.6 27B, designed to bring advanced AI capabilities to personal devices. This marks the first time a model of this size and complexity can run on a smartphone.

Bonsai 27B achieves this through innovative 1-bit and ternary weight quantization. The Ternary Bonsai 27B variant uses 1.71 effective bits per weight, resulting in a 5.9 GB model suitable for laptops, retaining full reasoning and tool-calling capabilities. The 1-bit Bonsai 27B variant, at just 3.9 GB, uses 1.125 effective bits per weight, fitting within the memory constraints of an iPhone 17 Pro. Both versions are multimodal, supporting vision tasks with a compact 4-bit vision tower.

Despite aggressive compression, Bonsai 27B retains significant intelligence, with the ternary version achieving 95% and the 1-bit version 90% of the full-precision baseline across a 15-benchmark suite. This high intelligence density is crucial for agentic workloads that require sustained, multi-step reasoning and tool use, enabling on-device AI assistants and workflows without the cost and privacy concerns of cloud-based solutions. The models are available under the Apache 2.0 License.