Startup Compresses 27-Billion-Parameter Model to Run on iPhone

PrismML, a startup that grew out of Caltech, announced that it has compressed the 27-billion-parameter Qwen 3.6 language model so it can run entirely locally on an iPhone 17 Pro, with no connection to cloud servers. The company says it achieved this without any drop in response quality, which it argues sets its approach apart from previous mobile compression techniques.

According to PrismML, the model handles extended conversations, multi-step reasoning, agentic tasks, and code generation, despite running entirely on the phone's chip with no internet access. That marks a significant departure from how smartphone makers have typically approached on-device AI, usually relying on smaller models or activating only part of them at a time.

How the compression works

Apple's current on-device models, run through the Core AI framework unveiled at WWDC 2026, rely on a sparse-activation architecture in which the model has roughly 20 billion parameters but only a small fraction, somewhere between 1 and 4 billion, are active at any given moment. PrismML took a different route, compressing the model's weights so aggressively that the entire model, all 27 billion parameters, fits in the phone's memory and stays fully active.

That approach is said to be more technically demanding, but in exchange it eliminates the quality trade-offs typical of sparse-activation models, where part of the model's knowledge is skipped for any given query. The company has not yet published full technical documentation of its compression method, saying it will do so alongside the model's open release in mid-July.

Talks with Apple

According to reports, Apple has held talks with PrismML about licensing its compression technology. That would matter a great deal to Apple, since the latest Siri features still rely in part on Google's Gemini model, an arrangement the Cupertino company, which champions privacy and on-device processing, reportedly views as a stopgap rather than a permanent solution.

Maybe in three years, 95 percent of the intelligence you need will be available locally, and only that last 5 percent will really require sending a query to the cloud - Babak Hassibi, CEO of PrismML

What it means for users and developers

For the average smartphone user, PrismML's breakthrough raises the prospect of a sophisticated AI assistant running fully offline, with no network latency, no conversation data sent to outside servers, and no cloud token fees. For mobile app developers, it signals that models once reserved for data centers could make their way directly into iOS apps within the next few months.

For Poland's mobile app market and startups building products on top of language models, this opens the door to designing AI features that run locally on the client's device, cutting cloud infrastructure costs and making it easier to meet personal data protection requirements that matter most in regulated sectors like banking and healthcare.

Sources: BigGo Finance (finance.biggo.com)