PrismML

What it does

PrismML has developed 1-Bit Bonsai, the first commercially viable large language models (LLMs) with 1-bit weights. These models use binary weights (e.g., -1, 0, 1) to drastically reduce memory footprint, latency, and energy consumption while maintaining competitive benchmark performance. The company also offers Ternary Bonsai (using {-1, 0, 1} weights) and Bonsai Image 4B for image generation. All models are available in sizes 8B, 4B, and 1.7B, and can be downloaded from Hugging Face.

Who it is for

PrismML targets developers and companies building applications for robotics, real-time agents, edge computing, and mobile devices. The models are designed to run on smartphones, GPUs, and other resource-constrained hardware. For example, Bonsai Image 4B is optimized for local inference on iPhone, Mac, and GPUs, and is available via the Bonsai Studio iOS app.

Why it matters

Large models are too big for smartphones and too energy-intensive for datacenters. PrismML's 1-bit approach delivers 14× less memory usage, 8× faster inference, and 5× less energy compared to full-precision counterparts. This enables AI capabilities on devices that previously could not run them, reducing cloud dependency and improving privacy. The company claims over 10× the intelligence density of full-precision equivalents.

Launch signal

PrismML launched on Hacker News as "Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs." The models are available for download on Hugging Face, and the Bonsai Studio app is on the iOS App Store. The company is supported by Khosla Ventures, Cerberus Capital, Google, and Caltech.

Brand and naming

The name "PrismML" suggests breaking down AI into fundamental components (like light through a prism), aligning with their focus on efficiency and density. "Bonsai" evokes the idea of a miniature yet fully functional tree, mirroring their goal of compressing large models into small, efficient forms. The brand positions itself as a research-driven alternative to the "bigger is better" trend in AI.