Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

Low-bit AI models that retain high performance become practical for devices and reduce cloud dependency.

PrismML has released Ternary Bonsai 2 27B, a highly compressed multimodal large language model. This new model, based on Qwen3.8 27B, utilizes ternary weights and group-wise scaling to achieve an effective 1.76 bits per weight, resulting in a 5.9GB footprint. Despite this significant reduction, Ternary Bonsai 2 27B retains 98.2% of its full-precision counterpart's benchmark performance across reasoning, coding, vision, and agentic tasks. The model supports a 262K-token context window and multimodal input, and is released under the Apache 2.0 license. It offers improved performance in areas critical for local applications like coding agents and multimodal workflows, with higher throughput and energy efficiency compared to previous versions and alternatives.

Ternary Bonsai 2 27B's achievement of near-lossless compression (98.2% performance retention at 1.76 bits per weight) signifies a major step forward in 'intelligence density'. This is crucial because it allows models with substantial capabilities, previously confined to powerful cloud infrastructure, to run efficiently on local hardware with significantly lower memory, compute, and power requirements. This breakthrough challenges the traditional trade-off between model size and performance, potentially democratizing access to advanced AI functionalities for edge devices and personal workstations.

The implications extend beyond local inference, impacting the economics and architecture of AI systems broadly. By fitting larger models into smaller envelopes, organizations can reduce serving costs, increase user capacity on existing hardware, and enable more sophisticated hybrid AI architectures. This efficiency gain is particularly relevant for AI-driven applications like coding assistants and multimodal agents, where faster iteration loops and continuous availability are paramount. The focus is shifting from raw model size to the most effective deployment of intelligence within budget constraints.