Startup Squeezes 27-Billion-Parameter Model into iPhone; Apple Already in Talks
Claire Weston
Startup PrismML compressed a 27-billion-parameter LLM from 54 GB to under 4 GB, running it locally on an iPhone; Apple has held talks on the technology, spotlighting a critical gap in its on-device AI strategy.
27 billion parameters on a phone — how does that work?
PrismML compressed Alibaba's open-source Qwen 3.6 model from roughly 54 GB to under 4 GB — a compression ratio above 90%.
The key difference: Apple's own on-device model has 200 billion parameters but uses a sparse architecture — a design that activates only a small slice at a time — so just 1–4 billion parameters fire per query. PrismML keeps all 27 billion parameters active simultaneously.
This means → running on the same phone, PrismML fields the entire team while Apple sends only a handful onto the pitch — in theory, the former handles more complex tasks.
Why does Apple need this technology?
Apple has long positioned on-device processing as the backbone of its privacy promise. Yet the Siri upgrade announced this June still relies on Google's Gemini model — its most advanced features run on Nvidia chips inside Google's cloud.
In plain terms = Apple says "everything stays on your phone," but the hardest work still goes to Google's servers.
Last year Apple tried to compress its own internal AI models to fit the iPhone and hit severe performance degradation — that is precisely why PrismML's technology carries strategic value for Cupertino.
Who is PrismML?
The company spun out of Caltech. CEO Babak Hassibi is a professor of electrical engineering there; the core compression patents are held by Caltech and exclusively licensed to PrismML.
PrismML closed a $16.25 million seed round earlier this year, with Khosla Ventures among the investors.
Khosla founder Vinod Khosla called PrismML a "fundamental breakthrough." The company says an open-source model will be released next Tuesday.
Can everything really run on a phone?
Not everyone buys the pure on-device path. Startups such as Argmax use a hybrid architecture — processing voice and image locally, then sending the data to the cloud for heavier reasoning.
The hybrid camp's core argument: cloud-hosted frontier models iterate at a pace of weekly updates; a purely on-device model cannot tap those gains.
This reflects a deeper divide — on-device AI optimizes for privacy and offline availability; cloud AI optimizes for continuous improvement. Neither side has settled the debate.
What comes next?
PrismML plans to compress even larger models — up to trillion-parameter scale — for on-device use, entering territory occupied by OpenAI's GPT and Anthropic's Claude.
Whether Apple ultimately moves to acquire or deeply partner with PrismML will be the clearest signal of how seriously it intends to close the on-device gap.
This means → if Apple acts, it concedes that its internal teams alone cannot solve the compression problem; if it does not, the timeline for true on-device AI likely stretches further.
Content is for reference only, not financial advice.