PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B That Run on Laptops and Phones

PrismML has announced the release of Bonsai 27B, a new iteration of the Qwen3.6-27B model. This release is characterized by the introduction of 1-bit and ternary builds, which allow the model to operate on consumer-grade hardware such as laptops and mobile phones. According to reports, the model is a full open reasoning model specifically engineered to fit within the hardware constraints of an iPhone. This development allows for advanced reasoning capabilities to be executed locally on a handheld device without relying on cloud infrastructure. Coverage of this release is appearing across several technology-focused outlets.

MarkTechPost explicitly identifies the technical specifications of the model, noting that it utilizes 1-bit and ternary builds of the Qwen3.6-27B architecture. The Decoder emphasizes the model's accessibility, highlighting that Bonsai 27B is a full open reasoning model capable of fitting on an iPhone. Meanwhile, Entrepreneur focuses on the scale of the technical achievement, reporting that the startup claims to have shrunk the AI model by 93% to achieve this portability. This release is significant because it represents a drastic reduction in the computational requirements for high-parameter models. The context provided by Entrepreneur suggests that this level of compression has attracted the attention of major industry players, stating that Apple wants to talk to the startup following these claims.

By reducing the model size by 93%, PrismML is positioning Bonsai 27B as a viable option for on-device AI, which typically requires significantly smaller models than the 27B parameter class to function efficiently on mobile operating systems. Future developments to watch include the outcome of the reported interest from Apple and the broader adoption of these 1-bit and ternary builds within the open-source community. Because the model is described as a full open reasoning model, developers may further iterate on its capabilities. Coverage does not yet specify the exact performance benchmarks compared to the original Qwen3.6-27B, but the focus remains on the ability to run these complex reasoning tasks locally on phones and laptops.