A Key Piece for Apple's On-Device AI? iPhone to Feature 27-Billion-Parameter Large Model for the First Time

Apple is seeking to keep more powerful AI capabilities on-device, and a startup backed by Khosla Ventures may have provided a key piece of the puzzle.

PrismML, a startup invested in by Khosla Ventures, claims to have successfully compressed a 27-billion-parameter large AI model to run locally on the iPhone 17 Pro, setting a new record for the scale of mobile AI models. The company stated that its compression technology incurs no performance loss, and the relevant open-source model will be officially released next Tuesday.

According to insiders, Apple has held discussions with PrismML regarding how to utilize its technology. Previously, The Information reported that Apple is actively seeking to acquire companies that can help it run more AI features on-device. Sources said that last year, Apple encountered significant performance degradation when attempting to compress its internal AI models to fit the iPhone.

27 Billion Parameters Fully Activated, Refreshing Mobile AI Record

PrismML stated that its compressed model is Qwen 3.6, an open-source large language model developed by Alibaba, with 27 billion parameters. In contrast, current mainstream mobile models have only billions of parameters activated at any given time.

The new on-device model released by Apple at its Worldwide Developers Conference in June has 20 billion parameters but uses a sparse architecture, with only 1 billion to 4 billion parameters activated at a time. PrismML's model keeps all 27 billion parameters simultaneously active during operation, a difference the company views as a core competitive advantage.

PrismML claims that the model is capable of handling complex conversations, reasoning, fully autonomous agents, and software programming tasks.

Mathematical Compression Technology Originates from Caltech, with Exclusive Patent License

PrismML is a spin-off from the California Institute of Technology (Caltech). Its CEO, Babak Hassibi, is a professor of electrical engineering at the university, and he completed the mathematical research underpinning the technology with his co-founders while they were students. Caltech holds the relevant patents and has exclusively licensed them to PrismML.

The company's core technology involves using a mathematical method to compress the size of the Qwen 3.6 model from approximately 54GB to less than 4GB, achieving a compression ratio of over 90%, with the company claiming no impact on performance.

PrismML completed a $16.25 million seed funding round earlier this year, with participation from Khosla Ventures. In an interview, Khosla Ventures founder Vinod Khosla stated that he was interested in PrismML because the company offered a "fundamental breakthrough." "When we invested in OpenAI in 2018, we heavily bet on the Transformer model, but what is the new way to build AI? Our team is always looking for new paths," he said.

Apple's On-Device AI Strategy and Potential Acquisition Logic

Apple has long made on-device AI a core pillar of its privacy and security commitments, largely avoiding the hundreds of billions of dollars in data center arms races undertaken by tech giants like Microsoft, Amazon, and Meta.

However, the long-awaited major Siri upgrade announced by Apple in June still relies on Google's Gemini model, with its most advanced features requiring Nvidia chips running on Google Cloud. This situation presents a clear gap with Apple's on-device AI vision, making PrismML's technology potentially strategically valuable to Apple.

Hassibi predicts that within the next three years, the vast majority of AI computing required by users will be performed locally. "Imagine that perhaps in three years, 95% of the intelligence you need will be available locally—on your phone, laptop, and home appliances—with only the final 5% of high-end needs truly requiring the cloud," he said. "I think this is the direction people see ahead."

Hybrid Architecture Proponents Pose a Challenge

Not all industry insiders agree with the pure on-device AI route. Startups like Argmax adopt a hybrid architecture, completing processing tasks such as speech and image recognition on-device before uploading information to the cloud for more complex reasoning.

Supporters of the hybrid architecture point out that cloud-based large models are currently iterating rapidly, with weekly updates. AI models running entirely on-device will struggle to benefit from the performance advantages of the latest and most advanced cloud models. This challenge is one of the core issues PrismML must continue to address in its commercialization path.

PrismML stated that the company plans to continue compressing larger-scale models—including those at the trillion-parameter level—to run on-device, at which point it will enter the arena competing with OpenAI's GPT and Anthropic's Claude.