PrismML has released Ternary Bonsai 2 27B, a new ternary-weight language model based on Qwen3.8 27B. According to reporting by MarkTechPost, the model requires 5.93 GB of storage, a significant reduction compared to the 53.80 GB footprint of the original FP16 version. PrismML states that the model retains 98.2% of the parent model’s average performance across a suite of 20 benchmarks. The architecture handles both text and image inputs and accommodates a context window of 262K tokens. PrismML has demonstrated the system running coding tasks through Cline.
Efficiency Through Ternary Quantization
Ternary weight quantization represents a technique in machine learning designed to drastically shrink model sizes and memory bandwidth requirements by reducing precision down to three distinct values. This approach allows large language models to run on significantly more modest hardware setups than their uncompressed counterparts usually demand, lowering deployment costs for developers who want to run capable vision-language systems locally or on edge devices without sacrificing the entirety of a model’s foundational capabilities.
Unanswered Questions and Future Hardware Needs
Despite the release details provided by PrismML, several technical specifications remain undisclosed. The announcement did not specify the exact hardware configurations, such as specific GPU models or memory speeds, required to run Ternary Bonsai 2 27B efficiently at its full 262K-token context length. Furthermore, the exact breakdown of performance retention across individual benchmarks was not detailed in the available reports, leaving developers to test the vision and text capabilities independently to verify how specific coding and multimodal workflows perform in production environments.
Developers interested in open-weight models can evaluate Ternuary Bonsai 2 27B under its Apache 2.0 license now.
Lead image: Jakub Pabis / Pexels
