PrismML released Ternary Bonsai 2 27B, a compressed version of the Qwen3.8 27B model that reduces the memory footprint to 5.9 GB. The AI model uses ternary weights to retain 98.2% of the aggregate benchmark performance of its full-precision counterpart and is available today under the Apache 2.0 license. It targets agentic workloads, with developers focusing on multimodal reasoning, agentic coding, and long-horizon tool use.

The model can run locally in a web browser via WebGPU or on consumer hardware such as the NVIDIA GeForce RTX 3060. Performance tests on an RTX 3060 with 12 GB VRAM showed a 220K context window and 35 tokens per second decode. This second iteration of the Bonsai line reduces silent failures and derailments during multi-step tasks compared to the initial 27B release.