PrismML releases Ternary Bonsai 2 27B: a 27B multimodal model in 5.9 GB
In one sentence On September 17 PrismML published Ternary Bonsai 2 27B, a ternary-weight version of Qwen3.8 27B: 1.76 effective bits per weight, 5.9 GB total, Apache 2.0 license, retaining 98.2% of the original's aggregate benchmark performance per the vendor.
A 27-billion-parameter language model normally takes tens of gigabytes and a high-end GPU. PrismML took Qwen3.8 27B and recompressed it with "ternary" weights: each weight can only be -1, 0 or +1, plus a scale factor shared by groups of weights. The result weighs 5.9 GB, more than nine times smaller than the original.
According to the company's published tests, the compressed model keeps 98.2% of the original's overall score (83.9 versus 85.4), and in some areas such as instruction following it scores slightly higher. It accepts text and images and has a 262K-token context.
The practical point is that it runs on user hardware: NVIDIA GPUs via CUDA and Apple devices (Mac, iPhone, iPad) via MLX, with dedicated kernels. The license is Apache 2.0, so commercial use is allowed.
Companies
PrismML
Tools
Ternary Bonsai 2 27B, Qwen3.8 27B, MLX, CUDA
Tags
Sources