Bonsai 27B is a full open reasoning model that fits on an iPhone

What changed

PrismML has compressed a 27-billion-parameter AI model known as Bonsai 27B to under 4 gigabytes. This shrinks it enough to run directly on an iPhone without offloading computation to the cloud. The smallest compressed version retains around 90 percent of the original model’s reasoning capabilities. Performance in key tasks like math and coding suffers very little, indicating a strong balance between size and capability.

Why builders should care

This compression breakthrough shifts what’s possible for on-device AI. Running a full reasoning model locally removes reliance on servers, cuts latency, and strengthens privacy by keeping data on the phone. Mobile developers can now embed larger, more capable AI models into apps without requiring expensive backend infrastructure. This also lowers operating costs and reduces the need for constant internet access.

Apple reportedly testing this technology signals a race to boost on-device AI power. Competitors in the phone and device markets will feel pressure to accelerate their own compression and optimization work to keep pace.

The practical takeaway

Builders focused on mobile and edge AI should explore Bonsai 27B’s approach to compression and efficient inference. Smaller, near-original-performance models enable smarter assistants, offline coding tools, and real-time reasoning in apps. This reduces dependency on cloud costs and mitigates user friction from lag or connectivity gaps.

The model’s size under 4 GB aligns with mainstream smartphone storage availability, so adoption isn’t stalled by device limitations. Companies working on AI-enabled apps now have a clearer path to integrating powerful, privacy-friendly models without sacrificing usability or functionality.

What to watch next

Monitor Apple’s moves around integrating such compressed models into iPhones and other devices, as it could reset expectations for on-device AI performance. Watch for other firms adopting or improving similar compression techniques to bring large open models to the edge. The evolution of tooling for local AI inference on phones will be crucial to track, especially to see how developers leverage this for real-world applications.

AI Quick Briefs Editorial Desk