What would it mean for the AI landscape if a 27-billion-parameter model could run entirely on your iPhone without a cloud connection? PrismML, a Khosla-backed startup led by CEO Babak Hassibi, has demonstrated exactly that capability, and it has confirmed to CNBC that Apple is evaluating its technology in early-stage talks that could reshape how the tech giant approaches on-device artificial intelligence.
The startup claims it can compress a 27-billion-parameter Qwen model from 54 gigabytes down to under 4 gigabytes. That is a 93 percent reduction in size, achieved without specialized hardware. The compressed model fits comfortably within the 8 gigabytes of memory available on current iPhones, opening the door to running frontier-class AI models locally on devices that billions of people carry in their pockets.
For Apple, which has long emphasized on-device processing for privacy and security reasons, the technology could be transformative. It would allow the company to deploy far more capable AI features without sending user data to cloud servers, addressing both privacy concerns and the latency that comes with cloud-based inference.
The Compression Breakthrough
PrismML core technology tackles one of the fundamental barriers to on-device AI: model size. Large language models require enormous amounts of memory, both for storage and for inference. The Qwen3.6-27B model that PrismML used in its demonstration normally requires a machine with at least 64 gigabytes of memory, effectively limiting it to desktop and server environments.
The startup approach achieves its 93 percent size reduction through techniques that the company has not fully disclosed publicly. What is known is that the resulting model retains most of its capabilities despite the aggressive compression. Hassibi confirmed to CNBC that the compressed models do lose a few percentage points in performance compared to the full-size versions. The smaller model is weaker on factual reasoning, mathematics, and coding tasks, though it remains functional for a wide range of applications.
This tradeoff between size and capability is the central engineering challenge in model compression. The question for Apple is whether the performance loss is acceptable for the kinds of tasks users perform on their phones. For conversational AI, summarization, and content generation, the compressed model may be more than adequate. For specialized reasoning tasks, the cloud remains a fallback option.
What Apple Is Evaluating
Hassibi told CNBC that Apple and other companies are evaluating PrismML technology, describing the talks as early and progressing nicely. He did not specify what form a potential deal might take. Industry observers have framed the range of possibilities as spanning from a straightforward technology licensing agreement to an outright acquisition of the startup.
Apple interest is consistent with its stated strategy around AI. The company has positioned privacy as a competitive advantage, and keeping AI processing on the device is central to that narrative. Currently, Apple Intelligence features rely on a combination of on-device models for simpler tasks and cloud-based inference through Private Cloud Compute for more complex requests. PrismML compression technology could shift that balance heavily toward on-device processing.
For Apple, the calculus goes beyond just technical capability. Running AI on the device eliminates the need for an internet connection, reduces operational costs associated with cloud inference, and sidesteps the regulatory scrutiny that comes with sending user data to external servers. If PrismML technology works at scale, it could be a key enabler for the next generation of iPhone AI features.
Market and Competitive Implications
PrismML emergence comes at a time when the entire smartphone industry is racing to bring more capable AI models on-device. Google has been running smaller versions of its Gemini models on Pixel devices. Samsung has partnered with Qualcomm to optimize AI inference on its flagship phones. Apple has lagged in on-device model capability relative to these competitors, making a potential PrismML partnership strategically significant.
The compression technology could also have implications beyond smartphones. Autonomous vehicles, Internet of Things devices, medical equipment, and edge computing infrastructure all face similar constraints around memory and compute. A 93 percent model compression technique that works across architectures would have broad applicability across industries.
Khosla Ventures involvement signals that the technology has passed early-stage scrutiny from experienced AI investors. The firm has backed multiple successful AI companies and tends to invest where it sees asymmetric technical advantage. The bet on PrismML appears to be that model compression, rather than model scale, will be the next frontier of AI value creation.
What This Means for Founders
PrismML trajectory offers several lessons for founders building in AI infrastructure. First, model compression and efficiency are increasingly becoming first-class problems alongside model capability. The market is rewarding companies that can solve the deployment problem, not just the training problem. Second, enterprise deals with large platform companies represent a viable go-to-market strategy for infrastructure startups. Even early-stage talks with a company like Apple can validate a technology and attract further investment.
For founders building AI-powered products, the implication is clear: on-device AI is about to become significantly more capable. If compression techniques like PrismML reach production readiness, the assumption that frontier models require cloud connectivity will no longer hold. Products that are designed for on-device inference from the start will have a structural advantage over those built on cloud dependency.
The talks between Apple and PrismML are still in their early stages, and no deal is guaranteed. But the fact that they are happening at all signals a recognition at the highest levels of the tech industry that model compression may be the key to unlocking the next phase of AI adoption.
