PrismML announced Bonsai 27B on July 14, 2026, marking the first 27-billion-parameter-class AI model capable of running on a phone. Based on Qwen3.6 27B, Bonsai 27B supports multi-step reasoning, structured tool calls, vision tasks, and agentic loops, all within a compact size suitable for mobile devices, according to prismml.com.
The model comes in two variants: Ternary Bonsai 27B, which uses ternary weights with FP16 group-wise scaling to achieve an effective 1.71 bits per weight at 5.9 GB, optimized for quality and capable of running on everyday laptops; and a second variant designed for phones. This development overcomes the previous barrier where 27B models required roughly 54GB in 16-bit precision, making them too large for phones or most laptops, prismml.com detailed.
Bonsai 27B extends the frontier of local AI deployment by enabling a capability tier previously impractical for mobile devices. Its features include coherent multi-step reasoning and vision tasks, which are critical for advanced AI applications on edge devices. This contrasts with earlier models that lacked the efficiency to operate on consumer hardware, positioning Bonsai 27B as a notable advancement in AI model compression and deployment.
PrismML’s announcement on July 14, 2026, establishes Bonsai 27B as a pioneering model in mobile AI, with the Ternary variant running on laptops at 5.9 GB. The company’s blog post on prismml.com provides further technical details and availability information.