PrismML unveils Bonsai 27B for iPhone as Apple reportedly tests its compression technology
By
Mr Bagel
Apple is in early discussions with PrismML, a startup backed by Khosla Ventures that claims to dramatically compress large AI models so they can run directly on an iPhone, according to multiple reports. The technology could slash memory usage by up to 15 times, potentially enabling more powerful on-device AI features while preserving user privacy.
PrismML has also released Bonsai 27B, which it bills as the first 27-billion-parameter model capable of running locally on a phone. Based on Qwen3.6 27B, the model uses 1-bit and ternary weight compression to fit on-device, supporting multimodal tasks like multi-step reasoning, structured tool calls, vision, and agentic computer-use loops, Hacker News reported.
"Apple is evaluating PrismML's technology, which the startup says can shrink powerful AI models enough to run directly on an iPhone while using up to 15x less memory."
The Cupertino giant's interest aligns with its longstanding privacy-first approach. Running AI models locally rather than in the cloud could improve both security and responsiveness, particularly for features like Siri. Crypto Briefing noted that Apple's exploration could "revolutionize on-device AI, enhancing privacy, reducing costs, and improving user experience."
"PrismML's local 27B iPhone model catalyst signals that the true strategic prize of edge AI is the OS-level inference router."
While the talks are still early, according to macdailynews.com, the release of Bonsai 27B gives Apple a concrete demonstration of what compressed models can achieve. PrismML's CEO told CNBC that Apple is looking at the company's technology, and 9to5Toys highlighted that the model runs on iPhone, iPad, and Mac, suggesting broad potential across Apple's ecosystem.
The reporting
7 outlets covered this story. Each links to the original.

Baker's Take
Comments
Sign in to join the conversation.
No comments yet. Be the first.