Apple’s on-device AI is missing a key piece? iPhone впервые packs a 27-billion-parameter large model for the first time

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

PrismML, a startup invested in by Khosla Ventures, claims to have successfully compressed a large AI model with 27 billion parameters to run locally on an iPhone 17 Pro, setting a new record for mobile AI model size. The company says its compression technology causes no performance loss, and the open-source model will be officially released next Tuesday.

According to sources familiar with the matter, Apple has held talks with PrismML about how to use its technology. The Information previously reported that Apple is actively seeking to acquire companies that can help it run more AI features on-device. Sources say that last year, Apple encountered significant performance degradation when trying to compress its internal AI models to fit the iPhone.

27 Billion Parameters Fully Activated, Setting a New Mobile AI Record

PrismML stated that the compressed model is Alibaba's open-source large language model Qwen 3.6, with 27 billion parameters. In contrast, current mainstream mobile models only have a few billion parameters activated at any time.

Apple's new on-device model, announced at WWDC in June, has 20 billion parameters but uses a sparse architecture, with only 1 to 4 billion parameters activated at a time. PrismML's model keeps all 27 billion parameters active simultaneously, a difference the company sees as a core competitive advantage.

PrismML says the model can handle complex conversations, reasoning, fully autonomous agents, and software programming tasks.

Mathematical Compression Technology Originates from Caltech, Exclusively Licensed Patent

PrismML is a spin-off from the California Institute of Technology (Caltech). Its CEO, Babak Hassibi, is a professor of electrical engineering at the school and completed the mathematical research underpinning the technology with co-founders during their time at Caltech. Caltech holds the relevant patents and has exclusively licensed them to PrismML.

The company's core technology uses a mathematical method to compress the volume of the Qwen 3.6 model from about 54GB to less than 4GB, a compression ratio exceeding 90%, and the company claims performance is unaffected.

PrismML completed a $16.25 million seed round earlier this year, with Khosla Ventures participating. Vinod Khosla, founder of Khosla Ventures, said in an interview that he was interested in PrismML because it offers a "fundamental breakthrough." "When we invested in OpenAI in 2018, we made a big 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 data center arms race that has cost Microsoft, Amazon, Meta and other tech giants hundreds of billions of dollars.

However, Apple's long-overdue major Siri upgrade announced in June still relies on Google's Gemini model, with its most advanced features needing to call on Nvidia chips running on Google Cloud. This gap between reality and Apple's on-device AI vision makes PrismML's technology potentially strategically valuable for Apple.

Hassibi predicts that within three years, the vast majority of AI computing users need will be done locally. "Imagine, perhaps three years from now, 95% of the intelligence you need can be obtained locally — on your phone, laptop, home appliances — and only the last 5% of high-end needs will actually have to go to the cloud," he said. "I think that's the direction people see."

Hybrid Architecture Proponents Raise Challenges

Not everyone in the industry agrees with the pure on-device AI approach. Startups like Argmax use a hybrid architecture, processing tasks like voice and images on-device before uploading information to the cloud for more complex reasoning.

Hybrid architecture supporters point out that cloud-based large models are currently being updated weekly, and AI models that run entirely on-device will struggle to benefit from the performance gains offered by the latest and most advanced cloud models. This challenge is one of the core issues PrismML will need to address on its path to commercialization.

PrismML says it plans to continue compressing even larger models — including trillion-parameter models — to run on-device, at which point it will enter a competitive arena with OpenAI GPT and Anthropic Claude.

Risk Warning and Disclaimer

        Market risk: Investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at your own risk.