Open Source AI

Ternary Bonsai 2: 27B Parameter Model Shrinks to Under 6GB

The release of Ternary Bonsai 2, a 27B parameter model derived from Qwen3.8, demonstrates a significant leap in model compression. By utilizing ternary weights, the model achieves a footprint of less than 6GB while reportedly retaining 98.2% of its original intelligence.

The AI Multiverse Editorial Desk•Thursday, September 17, 2026•2 min read

Reported viaReddit LocalLLaMA (Model Leaks & Local AI)

Reader:

Executive Brief & 3-Point Takeaways

Ternary Bonsai 2 achieves a <6GB footprint for a 27B parameter model.

The model retains 98.2% of the original Qwen3.8 intelligence through ternary weight compression.

WebGPU support enables local, browser-based inference for large-scale models.

Impact:

This development drastically lowers the hardware barrier for running high-parameter models, enabling complex AI inference directly within web browsers via WebGPU.

Fact-checked against primary sourcesCategory: Open Source AI

Ternary Bonsai 2: 27B Parameter Model Shrinks to Under 6GB

Illustration: AI-generated

Key Findings & Takeaways

Ternary Bonsai 2 achieves a <6GB footprint for a 27B parameter model.
The model retains 98.2% of the original Qwen3.8 intelligence through ternary weight compression.
WebGPU support enables local, browser-based inference for large-scale models.

Why It Matters

This development drastically lowers the hardware barrier for running high-parameter models, enabling complex AI inference directly within web browsers via WebGPU.

Key Details

Ternary Bonsai 2 has been released on Hugging Face, offering a highly compressed version of the 27B parameter Qwen3.8 model. The model is designed for extreme efficiency, allowing it to run locally in-browser using WebGPU kernels.

Technical Architecture

The model maintains the original hybrid-attention causal language model architecture of Qwen3.8. The primary innovation lies in the use of ternary weights, which allows for a 9x reduction in size compared to standard FP16 precision. Despite this aggressive compression, the developers report that the model retains 98.2% of the original model's performance capabilities.

What to Expect

Developers and enthusiasts can now access the model via the Prism ML collection on Hugging Face. A live demonstration is available through the webml-community Hugging Face Space, showcasing the feasibility of running large-scale models on consumer-grade hardware without the need for massive VRAM overhead.