Apple is in discussions with PrismML, a Khosla Ventures-backed startup that claims to have achieved a major breakthrough in running exceptionally large AI models directly on iPhones without relying on cloud servers, Aaron Tilley reports for The Information. The company says it has successfully compressed and executed a 27-billion-parameter model on the latest iPhone hardware, marking what it describes as the largest AI model ever run fully on-device.
PrismML, which emerged from stealth earlier this year as a Caltech spinout, specializes in extreme model compression techniques, particularly 1-bit and ternary weight architectures. These methods drastically reduce memory footprints while aiming to preserve performance. According to the startup, its technology allowed Alibaba’s open-source Qwen 3.6 (27B parameters) — which normally requires about 54 GB in standard precision — to shrink to under 4 GB. The compressed model reportedly ran with all parameters active simultaneously on an iPhone 17 Pro, enabling more complex on-device tasks like advanced reasoning and software development assistance.
This development aligns with Apple’s long-standing push to enhance Apple Intelligence with more powerful on-device capabilities. By keeping AI processing local, Apple can improve privacy, reduce latency, and lessen dependence on cloud infrastructure and external partners. Sources familiar with the matter indicate Apple has held talks with PrismML about potential integration of its compression technology to expand what future iPhones can handle natively.
The Technical Leap: 1-Bit Models and Intelligence Density
Traditional large language models use 16-bit floating-point precision, leading to massive memory requirements. PrismML’s approach uses 1-bit weights (essentially on/off states) and ternary variants (weights of -1, 0, or +1). The company’s flagship open-source “Bonsai” family includes:
• Bonsai 8B: An 8-billion-parameter model that fits in roughly 1 GB (versus ~16 GB for a typical 16-bit equivalent), while delivering competitive results on benchmarks like MMLU, GSM8K, and HumanEval.
• Smaller variants (4B and 1.7B) optimized for even greater efficiency.
PrismML claims these models offer up to 14× smaller memory use, 8× faster inference, and 5× better energy efficiency compared to full-precision counterparts, resulting in over 10× “intelligence density.” The startup has open-sourced its models under Apache 2.0 and released custom kernels for Apple’s Metal framework, making them compatible with iPhone and Mac hardware.
Co-founders, including Caltech professor Babak Hassibi and other PhDs, emphasize that the breakthrough stems from years of research into neural network compression. Investors include Khosla Ventures, Cerberus Capital, and Caltech itself, with the company raising $16.25 million in seed funding. Vinod Khosla has praised the work as a “mathematical breakthrough” that could shift AI from data-center dominance to efficient edge deployment.
Implications for Apple and the Industry
For Apple, adopting such technology could supercharge features in iOS and Apple Intelligence, enabling more sophisticated AI agents, real-time coding help, or advanced multimodal capabilities without constant server calls. It also addresses growing concerns around energy consumption and data privacy in AI.
PrismML positions its tech not just for consumer devices but for robotics, wearables, and industrial edge applications. By open-sourcing core models, the company hopes to accelerate adoption and invite community improvements.
While independent verification of all performance claims is still emerging, the reported iPhone demo and Apple’s interest signal strong validation. As the AI race intensifies, solutions that pack more intelligence into everyday devices — without ballooning costs or power draw — could prove transformative.
PrismML plans further releases and is expected to continue engaging with major device makers.
MacDailyNews Take: For Apple, which has invested heavily in on-device AI silicon like the Neural Engine, this could represent a key for the next-gen of truly personal, secure, and private artificial intelligence.
Please help support MacDailyNews — and enjoy subscriber-only articles, comments, chat, and more — by subscribing to our Substack: macdailynews.substack.com. Thank you!
Support MacDailyNews at no extra cost to you by using this link to shop at Amazon.
