PrismML has compressed a 27-billion-parameter AI model to under 4 GB, enabling it to run directly on iPhone devices. The compressed model retains 90 percent of its original performance with minimal impact on math and coding capabilities.
PrismML's Bonsai 27B achieves a significant milestone in mobile AI by fitting a full reasoning model on standard iPhones. The compression reduces the model from its original size to under 4 GB—small enough for on-device deployment without requiring cloud connectivity.
Benchmark results show the smallest version maintains 90 percent of the original model's performance across general tasks. More critically, math and coding scores remain largely unaffected, preserving the reasoning capabilities that distinguish advanced AI models from simpler alternatives.
The breakthrough addresses a key limitation of current mobile AI: most sophisticated models require cloud processing, introducing latency and privacy concerns. Running reasoning models locally on devices could enable faster inference and keep user data on-device.
Apple is reportedly already testing PrismML's compression technology, according to the source material. This suggests the iPhone maker sees practical value in deploying advanced reasoning capabilities to its devices. Apple has emphasized on-device AI processing as a privacy differentiator, making compressed reasoning models strategically important to its vision.
The achievement highlights ongoing progress in model compression techniques. As AI models grow more capable, reducing their computational footprint becomes increasingly valuable for consumer devices. Bonsai 27B demonstrates that aggressive compression doesn't necessarily sacrifice the advanced reasoning features users expect from modern AI systems.
Wider adoption of on-device reasoning models could shift how smartphones handle AI tasks, reducing dependence on cloud infrastructure and improving response times for latency-sensitive applications. However, the practical impact will depend on whether developers integrate these models into mainstream applications and whether the compressed versions prove sufficient for real-world use cases beyond benchmarks.
Israel-based Hemispheric secured $52 million in funding for its AI model that analyzes non-invasive brain activity measurements and converts them into quantitative diagnostic metrics.
Anthropic and Blackstone are backing Ode, a new venture that embeds AI engineers directly inside enterprises. The bet signals a shift in where the next trillion dollars in AI value may be created: not in building models, but in implementing them.
Spectro Cloud, an AI infrastructure company focused on managing token costs, secured $100 million in Series D funding at a valuation exceeding $1 billion. The raise marks significant growth from the company's $750 million valuation in 2024.
Startups like Altur are deploying AI chatbots to handle debt collection calls, automating a process traditionally done by humans. Y Combinator has backed six debt collection and settlement startups over the past six years.