Bonsai Image 4B

PrismML has released Bonsai Image 4B, a new set of image generation models designed to run on personal devices like laptops, phones, and tablets. Unlike most image tools that require expensive cloud servers, this technology allows users to create high-quality pictures directly on their own hardware. This release is significant because it brings powerful image generation capabilities to devices that were previously too small to handle such complex models. It is the first model in its size category to run directly on an iPhone.

Benefits

The main advantage of Bonsai Image 4B is its ability to compress large models into tiny packages without losing too much quality. The team created two versions of the model to suit different needs. The 1-bit version uses binary weights to achieve extreme compression, reducing the model size by over eight times compared to standard versions. This makes it perfect for devices with very limited memory. The Ternary version uses a slightly larger package but offers better visual quality and prompt accuracy. It reduces the size by more than six times while retaining 95% of the performance of the original full-precision model. Both versions are open source, meaning developers can use them freely under the Apache 2.0 license. The technology also ensures privacy since all processing happens locally on the user's device rather than sending data to the cloud.

Use Cases

This tool is ideal for anyone who wants to generate images without relying on internet connections or paying for cloud services. It works well on Apple devices like iPhones, iPads, and Macs using Apple Silicon chips. It also runs on standard computer graphics cards known as CUDA GPUs. Users can generate images on their phones while commuting or create variations of a design locally on a laptop without waiting for server responses. The speed is impressive, with a powerful Mac M4 Pro chip generating a standard image in about six seconds. This is much faster than the standard method on the same device. The technology is also being integrated into an iOS app called Bonsai Studio, which allows users to try the model directly on their phones.

Pricing

The models are free to use. Both the 1-bit and Ternary versions are released with open weights and code under the Apache 2.0 license. There are no fees for downloading or running the software on your own hardware.

Vibes

The public reception highlights the breakthrough nature of this release. The article notes that Bonsai Image 4B is the first image model in its parameter class to run directly on an iPhone. This achievement marks a new deployment regime for image generation, making capable outputs accessible on devices that were previously out of reach. Users and developers appreciate the practical local inference on hardware that was once considered too small for this class of model. The performance benchmarks show that the Ternary version retains 95% of the accuracy of the full-precision FLUX.2 Klein 4B model while being significantly smaller. This balance of quality and efficiency is seen as a major step forward for local AI tools.

Additional Information

Bonsai Image 4B was developed by PrismML, a team of researchers from Caltech. The company was founded with support from Khosla Ventures, Cerberus, and Google. The team has spent years working on compressing neural networks without sacrificing their reasoning ability. In addition to the models, PrismML is launching Bonsai Studio, an iOS app designed for trying Bonsai Image 4B directly on an iPhone. The technology utilizes specific low-bit paths on Apple Silicon and Gemlite low-bit GEMM kernels on CUDA GPUs to achieve its efficiency. The compression allows these models to run on-device where full-precision versions cannot fit in memory.