A startup emerging from stealth today unveiled the world's first commercially viable 1-bit large language models (LLMs), fundamentally changing how artificial intelligence can be deployed on everyday devices. PrismML's flagship model, called 1-bit Bonsai 8B, delivers cutting-edge AI capabilities while operating efficiently on smartphones, laptops, and embedded systems, compared to traditional models that require massive datacenter infrastructure .

What Makes 1-Bit AI Different From Everything Else?

Most AI models today use 16-bit or 32-bit precision, meaning each parameter in the neural network requires significant memory and computing power. PrismML rethought this from the ground up by creating models with native 1-bit structure, where each parameter uses just a single bit of data instead of 16 or 32. This radical simplification doesn't sacrifice reasoning performance, according to the company's benchmarks .

The efficiency gains are striking. The 1-bit Bonsai 8B model is 14 times smaller, 8 times faster, and 4 to 5 times more energy efficient than leading full-precision 8-billion parameter models like Llama 3 8B, while maintaining competitive performance on intelligence benchmarks . In practical terms, the model requires just 1 gigabyte of memory compared to 16 gigabytes for traditional versions, making it feasible to run on consumer-grade hardware.

"We spent years developing the mathematical theory required to compress a neural network without losing its reasoning capabilities. We see 1-bit not as an endpoint, but as a starting point. We are creating a new paradigm for AI: one that adapts to diverse hardware environments and delivers maximum intelligence per unit of compute and energy," said Babak Hassibi, CEO and Founder of PrismML and Professor at Caltech.

Babak Hassibi, CEO and Founder of PrismML and Professor at Caltech

How to Deploy 1-Bit AI Models for Your Use Case

• Download and Test Locally: Developers can download the 1-bit Bonsai models under the Apache 2.0 open-source license for free starting today, enabling immediate experimentation on personal devices without cloud infrastructure .
• Choose the Right Model Size: PrismML is releasing three versions: the 1-bit Bonsai 8B with 1GB memory footprint, the 4B model with 0.5GB, and the 1.7B model with 0.24GB, allowing developers to match model capacity to their hardware constraints .
• Integrate Into Existing Workflows: The models are designed for seamless integration with existing AI workflows and are optimized for low-latency inference on consumer-grade CPUs, NPUs (neural processing units), and edge GPUs, meaning minimal code changes are needed .
• Build Edge-First Applications: The efficiency enables developers to create sophisticated AI applications that execute directly on devices, unlocking new possibilities in robotics, wearables, and personal computing that were previously impractical due to power and memory constraints .

Why This Breakthrough Matters Beyond Your Phone

While the immediate impact is enabling AI on edge devices, the implications extend far beyond consumer hardware. The same efficiency gains that allow local deployment also allow datacenters to operate more effectively by improving hardware utilization, lowering operating costs, and reducing energy consumption . As AI models grow larger and more computationally intensive, power consumption has become a critical bottleneck for scaling AI infrastructure.

"Power has become the ultimate bottleneck for scaling AI datacenters, and PrismML is fundamentally transforming the power-to-compute equation. Moreover, by reducing the memory footprint and bandwidth demands, this breakthrough technology has the potential to do more than just improve the economics of AI infrastructure; it can unlock a new frontier for innovation in computer architecture for AI inference and the next generation of AI models," noted Amir Salek of Cerberus Ventures, who also founded and led the TPU (tensor processing unit) program at Google.

Amir Salek, Cerberus Ventures and former TPU program lead at Google

The technology could reshape how companies design AI hardware itself. By reducing memory footprint and bandwidth demands, 1-bit models change the optimization equation for entire systems, from individual devices all the way up to cloud infrastructure . This has implications for how future AI chips and processors are engineered.

PrismML's breakthrough is built on proprietary intellectual property developed at Caltech and is backed by prominent venture capital firms including Khosla Ventures and Cerberus Ventures, along with compute grants from Google and Caltech . The company trained its models using Google v4 TPUs, demonstrating that the approach works at scale with modern AI infrastructure.

"AI's future will not be defined by who can build the largest datacenters. It will be defined by who can deliver the most intelligence per unit of energy and cost. PrismML represents that kind of breakthrough," said Vinod Khosla, Founder of Khosla Ventures and an investor in the company.

Vinod Khosla, Founder of Khosla Ventures

The launch represents a shift in how the AI industry thinks about model deployment. Rather than assuming all advanced AI must run in the cloud, PrismML demonstrates that with the right mathematical approach, powerful reasoning and language understanding can operate on constrained devices. This could accelerate the development of privacy-preserving AI applications, reduce latency for real-time use cases, and lower the barrier to entry for developers building AI-powered products.