PrismML Launches Bonsai 2, a 27B-Class Ternary Model Compressed 9x
Based on Qwen3.8 27B, PrismML has launched Bonsai 2 27B, a new flagship Ternary model compressed to 5.9GB with minimal performance loss.
Image source: https://x.com/PrismML/status/2100692248480596348
PrismML announced the release of "Bonsai 2 27B," a new flagship Ternary (3-bit/ternary-weight) model, on September 17, 2026, aimed at maximizing the efficiency of large language models (LLMs).
Bonsai 2 27B is built upon the Qwen3.8 27B architecture. By leveraging PrismML's proprietary ternary compression technology, the model size has been reduced by a factor of 9 compared to its original precision, resulting in a total model size of approximately 5.9GB.
Optimized Performance for On-Device AI
The core strength of Bonsai 2 lies in its ability to maintain high performance despite extreme compression. PrismML reports that the model retains 98.2% of its original comprehensive benchmark scores, enabling the capabilities of a large-scale model to be deployed in much lighter, on-device environments.
Key features include:
- Integrated Multimodality: Combines reasoning, coding, and vision capabilities into a single model.
- Resource Efficiency: The 5.9GB footprint allows for practical inference speeds on local hardware, reducing reliance on high-end server environments.
- Accessibility: The model is currently available via Hugging Face and PrismML Studio.
The ternary compression technology utilized in Bonsai 2 27B offers a compelling new alternative for practitioners aiming to deploy high-performance multimodal models on hardware-constrained on-device AI devices. Note that fully utilizing the potential of this ternary model may require optimized inference libraries or specialized hardware support.
Sources
PrismML, Bonsai 2 27B Launch Announcement - PRNewswire PrismML Announcement - X