{"slug": "apple-evaluates-prismml-for-ai-model-compression-on-iphones", "title": "Apple evaluates PrismML for AI model compression on iPhones", "summary": "Apple has held exploratory discussions with Caltech-born startup PrismML about integrating its AI model compression technology into iPhones. PrismML claims it can shrink a 54 GB AI model to under 4 GB using extreme quantization, enabling on-device AI that addresses privacy, cost, and latency concerns. No formal agreement has been reached, but the talks signal Apple's continued push for on-device AI processing.", "body_md": "# Apple evaluates PrismML for AI model compression on iPhones\n\nA Caltech-born startup claims it can shrink a 54 GB AI model to under 4 GB, and Apple is paying attention\n\nRunning a serious AI model on your phone has, until now, been a bit like trying to fit a grand piano through a cat flap. But a startup called PrismML appears to have found a way to make the piano smaller without losing too many keys, and Apple is reportedly sitting across the table from them trying to figure out what a deal might look like.\n\nAccording to a report from The Information, Apple has held discussions with PrismML about integrating the startup’s model compression technology into iPhones. The conversations are exploratory, with no formal agreement or deployment timeline confirmed.\n\n## Shrinking a 54 GB model to fit in your pocket\n\nHere’s what PrismML actually did. They took Alibaba’s Qwen 3.6, a 27 billion parameter AI model that weighs in at roughly 54 GB, and compressed it to under 4 GB. In English: they took something that would barely fit on most laptops and made it run seamlessly on an iPhone 17 Pro.\n\nThe company’s approach relies on extreme quantization, a technique that reduces the precision of the numbers used to represent a neural network’s weights. PrismML pushes this to the logical extreme, producing commercially viable 1-bit models. Their Bonsai 27B 1-bit variant clocks in at approximately 3.9 GB, while a ternary version sits around 5.9 GB and targets laptops.\n\nThe claimed improvements span memory usage, inference speed, and energy efficiency, all of which matter enormously when you’re working with a device that runs on a battery and fits in a jeans pocket.\n\nPrismML emerged from stealth on March 31, 2026, armed with $16.25 million in seed funding from Khosla Ventures, Cerberus Ventures, and Caltech. The company was founded in 2025, and its CEO and co-founder, Babak Hassibi, is a Caltech professor whose work on the mathematical foundations of neural network compression forms the backbone of the technology.\n\n## Why Apple cares about on-device AI\n\nApple’s interest here fits a pattern that’s been building for years. The company has consistently prioritized on-device processing for privacy-sensitive features, from Face ID to Siri’s speech recognition. Running AI models locally, rather than bouncing queries to cloud servers, addresses three problems Apple cares deeply about.\n\nFirst, privacy. Data that never leaves the device can’t be intercepted, subpoenaed, or leaked from a server farm. Second, cost. Cloud inference at scale is expensive. Every query to a large language model hosted on remote GPUs costs money. Multiply that by hundreds of millions of iPhones, and the economics get uncomfortable fast. Third, latency. Local inference is faster than a round trip to a data center. For real-time applications like voice assistants, augmented reality, or camera features, the difference between 50 milliseconds and 500 milliseconds is the difference between magic and frustration.\n\nThe challenge has always been that the most capable AI models are simply too large to run on mobile hardware. A 54 GB model is a non-starter on a phone with 8 GB of RAM. PrismML’s compression approach, if it works as advertised, removes that constraint.\n\n## What this means for investors and the broader market\n\nApple evaluating a startup is not the same as Apple acquiring or partnering with one. No deal has been announced, and it’s entirely possible that Apple’s own ML teams are working on competitive approaches.\n\nThe $16.25 million seed round, while modest by Big Tech standards, positions PrismML in a growing cohort of startups attacking the inference cost problem from different angles. The fact that Khosla Ventures led the funding suggests institutional conviction that model compression is a durable market.\n\nInvestors should watch for two things in the coming months. First, whether Apple makes any formal announcement about on-device model capabilities at future product events. Second, whether PrismML’s benchmarks hold up under independent scrutiny, because compressing a model by 93% and claiming it still works well is a bold assertion that needs rigorous third-party validation before anyone should price it into their portfolio decisions.\n\n**Disclosure:** This article was edited by Editorial Team. For more information on how we create and review content, see our\n\n[Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/apple-evaluates-prismml-for-ai-model-compression-on-iphones", "canonical_source": "https://cryptobriefing.com/apple-evaluates-prismml-ai-compression-iphones/", "published_at": "2026-07-14 19:48:52+00:00", "updated_at": "2026-07-14 20:03:36.957607+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-startups", "ai-products", "ai-infrastructure", "ai-research"], "entities": ["Apple", "PrismML", "Alibaba", "Qwen 3.6", "Khosla Ventures", "Cerberus Ventures", "Caltech", "Babak Hassibi"], "alternates": {"html": "https://wpnews.pro/news/apple-evaluates-prismml-for-ai-model-compression-on-iphones", "markdown": "https://wpnews.pro/news/apple-evaluates-prismml-for-ai-model-compression-on-iphones.md", "text": "https://wpnews.pro/news/apple-evaluates-prismml-for-ai-model-compression-on-iphones.txt", "jsonld": "https://wpnews.pro/news/apple-evaluates-prismml-for-ai-model-compression-on-iphones.jsonld"}}