Local LLMs Get Serious: Bonsai 27B Weighs Less Than 4 GB

The PrismML team has introduced the Bonsai 27B model series, which, thanks to 1-bit quantization, has been reduced in size from 54 GB to less than 4 GB while maintaining high quality.

Compiled by Sergey KostenchukPublished 2026-07-14Updated 2026-07-15

2026-07-14 Research HuggingFace

🤖 Local LLMs Get Serious: Bonsai 27B Weighs Less Than 4 GB

The PrismML team has introduced the Bonsai 27B model series, which has undergone extreme compression via 1-bit quantization. Thanks to this, the model weight has been reduced from the original ~54 GB to 3.8–3.9 GB while maintaining approximately 90% quality. A more advanced version, Ternary Bonsai 27B, weighs 5.9 GB and provides up to 95% of the base model's quality.

🌍 This breakthrough in extreme (1-bit) quantization proves the possibility of running powerful LLMs on memory-constrained devices, shifting the development focus from cloud computing to Edge AI.

👤 Now, high-class models at the 27B parameter level can be run locally on standard laptops, smartphones, or even directly in the browser via WebGPU, without sending data to external servers.

Source 1: https://huggingface.co/collections/prism-ml/bonsai-27b Source 2: https://huggingface.co/spaces/webml-community/bonsai-webgpu-kernels

Sources