A little-known startup called PrismML claims to have achieved a feat that could reshape how artificial intelligence works on smartphones. The company says it compressed Alibaba's open-source Qwen 3.6 — a large language model with 27 billion parameters — from roughly 54 gigabytes down to under 4 gigabytes, and successfully ran it on an iPhone 17 Pro. According to a report from The Information, Apple has already held meetings with PrismML to explore potential on-device uses for the technology.
What PrismML's Compression Means
Large language models like Qwen 3.6 are typically too big to fit entirely on a phone's memory. They usually run on powerful cloud servers, with users sending prompts over the internet and receiving answers back. PrismML's claim is significant because it suggests that the model can now live entirely on the device itself. That shift — from cloud-based inference to on-device inference — could bring several practical benefits for everyday users.
When AI processing happens locally, response times are faster because there is no round trip to a remote server. Privacy improves because your data never leaves the phone. And features can keep working even with weak or no internet connection. For Apple, which has increasingly emphasized on-device processing for privacy and performance, a model that fits within a phone's memory constraints could open the door to more advanced AI features without relying on cloud infrastructure.
Apple's Interest and the Bigger Picture
Apple's reported meetings with PrismML suggest the company is actively looking for ways to bring more capable AI models to its devices. Apple has long positioned privacy as a key differentiator, and on-device AI aligns with that strategy. The company already uses smaller neural engines for tasks like photo editing and voice recognition, but a model of Qwen's size could enable more sophisticated applications — such as advanced language understanding, summarization, or even code generation — all without sending data to the cloud.
PrismML's approach is part of a broader trend in the AI industry. As models grow larger and more capable, there is increasing pressure to make them efficient enough to run on consumer hardware. Other startups and research labs are also working on compression techniques, including quantization, pruning, and distillation. If PrismML's method proves scalable, it could give Apple and other smartphone makers a way to offer cutting-edge AI features without the cost and complexity of cloud servers.
The news also highlights the growing role of Chinese AI companies like Alibaba in the open-source ecosystem. Alibaba's Qwen models have become popular among developers because they are freely available and competitive with proprietary alternatives. However, China is considering new rules that could restrict foreign access to its AI models, which could affect how widely models like Qwen are adopted outside the country.
What It Means for Investors
For investors, the development touches on several themes. First, it underscores the importance of efficiency in AI. While much of the market's attention has been on training ever-larger models, the ability to run them on edge devices — phones, laptops, cars — is becoming a competitive advantage. Companies that can deliver powerful AI without requiring constant cloud connectivity may be better positioned in the consumer market.
Second, Apple's interest in PrismML is a reminder that the company is investing heavily in AI, even if it has been less vocal about it than some rivals. Apple's approach has historically been to integrate new technology into its ecosystem in a polished, user-friendly way. If on-device AI becomes a key feature of future iPhones, it could drive upgrade cycles and strengthen the company's services revenue.
Third, the news adds to the narrative around Chinese AI companies. Alibaba's Qwen models are gaining traction globally, and the company's stock has seen strong moves — Asian ADRs rose recently as Alibaba surged 9.8%. However, regulatory risks remain, and investors should watch how China's evolving rules on AI exports and foreign access play out.
Finally, the broader AI chip and infrastructure market could be affected. If more AI workloads move to devices, demand for cloud-based inference chips might moderate, while demand for efficient edge processors could rise. Startups like Positron and SambaNova are already raising large sums to build specialized hardware, and the shift toward on-device AI could accelerate that trend.
For now, PrismML's claim remains just that — a claim. The company has not released detailed benchmarks or independent verification. But the fact that Apple is willing to talk suggests the technology is worth watching. If it works at scale, the way we interact with AI on our phones could look very different in the next few years.
