Xenova shared the newly released Bonsai 27B large model from the Prism ML team on the X platform. This 27-billion-parameter model has rewritten the minimum size record for compressed on-device large models.

Originally, the 27B model took up 54GB of storage at FP16 precision. After 1-bit quantization compression, its size dropped directly to 3.8GB, achieving a compression rate of 93%. Official test data shows that the compressed model still retains 90% of the original model's capability.

More interestingly, paired with custom WebGPU kernels written with Fable 5 and GPT 5.6 Sol, this 27-billion-parameter large model requires no cloud server and runs directly in a local browser.

The entire model has now been open-sourced on Hugging Face, and you can try the demo directly online:

- [Hugging Face Model Collection Page](https://huggingface.co/collections/prism-ml/bonsai-27b)

- [Online Demo Experience](https://huggingface.co/spaces/webml-community/bonsai-webgpu-kernels)


After this achievement was announced, it blew up the tech community. Many industry practitioners marveled at how extreme this compression ratio is. Some joked that this is practically sorcery, while others noted that with the continuous evolution of WebGPU, browsers have become a legitimate AI runtime platform. Chinese developers also commented that running a 27B model natively in the browser really goes beyond conventional expectations.

But controversies soon followed. Some users directly questioned that 1-bit quantization inherently leads to precision loss, and the claim that it retains 90% of the original capability cannot be substantiated. 1-bit models simply cannot support serious tasks and can only handle simple conversations. Other users who tested the model reported that it failed basic capability tests, though they acknowledged the exploratory value of this research direction.

After testing the model, one user shared a recording of their actual operation:

Setting aside the precision debate, this achievement sends a very clear signal: the progress of on-device large model compression is much faster than most people expected. Just a few years ago, no one would have thought that a 27B large model could fit into less than 4GB of storage and run in a browser on an ordinary consumer device.

Some people have already asked whether the same method can be used to compress Gemma 4 30B. After all, this model has strong multilingual capabilities and good tool calling performance. If it could really be compressed to this size, average users could deploy it natively on their own devices.

You can go to the Hugging Face demo page to test it yourself and experience how this 3.8GB 27B large model actually performs in practice.