Tech
Mobile" revolution in artificial intelligence: Giant model that can run on iPhone"
- The company argues that the new approach could usher in an era of "intelligence density" in artificial intelligence.
- US-based AI startup PrismML aims to deliver higher performance with the same hardware by focusing on the concept of "intelligence density" rather than increasing the model size.
- Unlike traditional quantization methods that only shrink certain layers, PrismML transforms the entire model into a 1-bit or ternary weight structure.
US PrismML has found a new way to shrink artificial intelligence models. The company argues that the new approach could usher in an era of "intelligence density" in artificial intelligence. See Full Size Although artificial intelligence models continue to grow in line with the scaling law, this approach creates serious hardware and energy requirements, especially in devices such as phones, computers and robots.
US-based AI startup PrismML aims to deliver higher performance with the same hardware by focusing on the concept of "intelligence density" rather than increasing the model size. Unlike traditional quantization methods that only shrink certain layers, PrismML transforms the entire model into a 1-bit or ternary weight structure. The company also develops special inference kernels for this structure, enabling direct use of compressed weights.
Thus, model size, memory usage and data transfer are significantly reduced. See Full Size The 1-bit Bonsai 8B, introduced in March 2026 and using this technique, was the company's first model suitable for commercial use. According to PrismML, the model is 14 times smaller than the Llama 3 8B, works 8 times faster and 4-5 times more efficiently.
Source: Donanımhaber