Bonsai 27B is a full open reasoning model that fits on an iPhone
PrismML has released Bonsai, a compression technique that shrinks a 27-billion-parameter model to under 4GB, allowing it to run on an iPhone while retaining 90% of its performance.
- Bonsai compresses 27B models to under 4GB for mobile use.
- Benchmarks show 90% performance retention in math and coding.
- Apple is reportedly testing the technology for on-device AI.
PrismML introduced a compression method named Bonsai that reduces a 27-billion-parameter model to under 4GB. This size allows the model to operate locally on an iPhone, a device typically limited to much smaller models due to memory constraints.
According to internal benchmarks, the compressed version maintains 90% of the original model's capabilities. The company reports that performance in math and coding tasks remains largely unaffected despite the drastic reduction in size.
The technology has reportedly attracted the attention of Apple, which is currently testing the compression method. If adopted, this could significantly enhance the performance of on-device AI features on future Apple hardware.
Enables running larger, smarter models on consumer edge devices.
Offers a path to powerful, private, offline AI applications on mobile.
Signals a major shift towards capable AI running without the cloud.
LLMApple opens its new Siri AI to everyone with the iOS 27 public beta
LLMBonsai 27B: 1-bit dense LLM running locally in your browser using custom WebGPU kernels
LLMOpus 5 Could Be Out As Soon As This Week Per Reliable Leaker On Twitter
LLMKimi K3 in the next few hours. Deepseek V4 GA later in the week. New Liquid models. New Mistral models sometime this month. And some rumours suggest GLM 5.5 is coming in August. Openweight AI is eating good.
Google Delays Gemini 3.5 Pro Over Early Performance Flaws - Geeky Gadgets
SecurityNew Mac malware masquerades as Apple's crash reporter: 3 ways to dodge the threat
A new MacOS malware called CrashStealer is targeting user data, passwords, and cryptocurrency by masquerading as Apple's crash reporter. Users can protect themselves with precautions.
Amplifying impact: A bold commitment to responsible AI - The University of Utah
The University of Utah has made a commitment to responsible AI practices, aiming to amplify the positive impact of AI research.
DeepSeek weighs IPO as Zhipu embraces AGI - Semafor
Chinese AI lab DeepSeek is reportedly exploring an initial public offering. Meanwhile, competitor Zhipu AI is shifting its strategic focus toward achieving artificial general intelligence.
BusinessHack suggests AI music generator Suno scraped YouTube for training data
A security breach at AI music generator Suno has reportedly revealed that the company scraped decades of audio, potentially including content from YouTube, for its training data.
BusinessWhatnot acquires Shaped to power real-time live shopping recommendations
Whatnot acquires AI startup Shaped to enhance real-time live shopping recommendations and personalization.
Affordable alternatives to Nvidia servers crop up for Japan AI players - Nikkei Asia
Nvidia server alternatives emerge in Japan, offering cost-effective solutions for AI players.