Breaking Cloud Dependency: PrismML Unveils 'Bonsai', a 27B AI Model Capable of Running on Smartphones

On July 14, 2026, AI startup PrismML officially announced 'Bonsai 27B', the first 27-billion parameter AI model capable of running natively on smartphones. This announcement marks a breakthrough event that brings cloud-based large language models inside mobile devices, once again pushing the performance boundaries of on-device AI.

Bonsai 27B maintains powerful reasoning capabilities while compressing the model size from the original 54GB down to 3.9GB. Industry experts are calling this the 'DeepSeek moment' of edge computing, suggesting the dawn of an era where high-performance AI services can be used without a data center connection.

PrismML's achievement demonstrates a leap forward in model compression technology. The fact that a model with 27 billion parameters can run smoothly on consumer hardware like the iPhone 17 Pro is expected to accelerate the shift from the existing cloud-first paradigm to a focus on local intelligence.

Bonsai 27B delivers large-model-class intelligence, previously dependent on the cloud, directly into the user's pocket, marking one of the most significant milestones in the history of on-device AI.

The technical core lies in the 1-bit and ternary quantization architecture independently developed by PrismML. This technology succeeded in preserving the intelligent structure required for complex reasoning tasks while reducing memory usage by more than 14 times by extremely simplifying the model's weights.

Technical Architecture: 1-bit and Ternary Quantization

Bonsai 27B is divided into a 1-bit variant that represents weights as {-1, +1} and a ternary variant that represents weights as {-1, 0, +1}. In particular, by introducing the 'Binary g128' format and assigning one FP16 scale factor for every 128-weight group, it achieved efficiency by reducing the effective number of bits per weight to 1.125 bits.

• ['1비트 변체: 3.9GB 메모리 점유, 아이폰 17 프로 등 모바일 기기 최적화.', '삼진 변체: 5.9GB 메모리 점유, 노트북 및 고사양 모바일 기기용.', '기반 모델: Qwen3.6-27B의 아키텍처를 유지하면서 저비트 표현으로 변환.']

In terms of performance, Bonsai 27B shows remarkable efficiency. The 1-bit variant was found to maintain approximately 90% of the performance of the full-precision model in 15 major benchmark tests, which means it provides a practical level of intelligence in a local environment beyond simple compression.

The emergence of such high-performance local models also contributes to protecting user privacy. Since all data processing takes place inside the device, sensitive information does not need to be transmitted to external servers, and consistent AI performance can be guaranteed even in environments with unstable network connections.

Hardware Requirements and Developer Ecosystem

PrismML is encouraging participation from the open-source community by releasing Bonsai 27B under the Apache 2.0 license. In communities such as 'LocalLLaMA' on Reddit, local operation test results for Bonsai 27B are already being shared, and it is receiving favorable reviews for generating answers of similar quality with much less memory than existing 8-bit quantized models.

On the hardware side, it actively utilizes the NPU and GPU resources of the latest mobile devices such as the iPhone 17 Pro. The 3.9GB memory footprint is well within the available RAM range of current premium smartphones, allowing users to experience large-model-class AI without separate hardware upgrades.

Finally, this technical breakthrough is expected to influence the strategies of global big tech companies. According to foreign media outlets such as CNBC, news has emerged that Apple is considering the adoption of PrismML's technology, which is expected to further strengthen the trend of integrating high-performance AI as a standard feature in future smartphones.