Apple Inc. is in early talks with PrismML, a Silicon Valley startup that claims it can radically shrink powerful artificial intelligence (AI) models to run directly on consumer devices, including the iPhone.
The Khosla Ventures-backed startup, a spinoff from the California Institute of Technology (Caltech), recently released a compressed version of Alibaba’s open-source Qwen model. PrismML successfully shrunk the model from roughly 54 gigabytes to under 4 gigabytes, allowing all 27 billion of its parameters to run locally on an iPhone 15 or newer.
PrismML CEO Babak Hassibi confirmed that Apple and other tech companies are actively evaluating the technology’s speed, energy efficiency, and on-device performance. While Hassibi characterized the discussions as early, he noted that “things are progressing nicely.”
Apple has not yet commented on the talks.
The negotiations coincide with Apple’s public beta release of iOS 27, which introduces a long-anticipated overhaul of Siri. As Apple seeks to compete with rival assistants from OpenAI and Anthropic, running AI directly on-device is critical to its strategy.
Local processing offers several key advantages. It keeps sensitive personal, health, and fitness data off external cloud servers. It eliminates the delays associated with sending data to remote networks. And it lowers cloud-computing expenses and enables offline functionality.
The shift is also financially urgent. Morgan Stanley estimates that Apple’s average DRAM and NAND memory costs could surge by 190% and 180% year-over-year in fiscal 2027, potentially forcing a $200 price hike on upcoming iPhone 18 models. Keeping models small helps Apple fit advanced AI into tighter hardware limits without forcing costly memory upgrades.
PrismML achieves dramatic compression by simplifying how data is stored, reducing standard 16-bit values to just one or three possible values. According to the company, this method uses 10 to 15 times less memory, generates responses six to eight times faster, and consumes three to six times less energy than conventional models.
However, Hassibi acknowledged a trade-off: Compressed models typically lose a few percentage points of overall performance, with factual recall declining slightly before reasoning, math, and coding skills are affected.
While the technology promises to run server-grade models on everyday laptops and phones, analysts urge caution.
Experts from Counterpoint Research and IDC warn that PrismML’s claims must be proven at scale. The ultimate test will be how the technology handles millions of concurrent queries, lengthy prompts, and background battery drain during multitasking.
Experts also debate whether extreme compression will cool the AI hardware market. While PrismML’s technology could allow a model that normally requires eight GPUs to run on just one, analysts suggest it may not lower overall chip demand. Instead, it could simply shift processor and memory demand from massive cloud datacenters directly into consumer edge devices.
For now, PrismML is offering two compressed versions of its model for free. The startup plans to compress Google’s Gemma model next, followed by even larger frontier models, potentially expanding local AI capabilities to robotics and autonomous systems.
