PrismML releases Ternary Bonsai 2 27B, a 5.9GB multimodal model retaining 98.2% benchmark performance

PrismML introduced Ternary Bonsai 2 27B, a compressed model based on Qwen3.8 27B that uses ternary weights and FP16 group-wise scaling for a 5.9GB footprint. The company says the model is over 9x smaller than its full-precision counterpart while retaining 98.2% of aggregate benchmark performance, with support for a 262K-token context window, multimodal inputs, and an Apache 2.0 license.

The release highlights how aggressive model compression is becoming a practical path to running larger AI systems locally without major capability loss. If the performance claims hold in real-world use, models like Bonsai 2 could broaden on-device AI deployment across laptops, edge devices, and privacy-sensitive enterprise environments.