Two months after introducing its first 27-billion-parameter multimodal model, the company has unveiled Ternary Bonsai 2 27B, a compressed version that retains 98.2% of the performance of its full-precision counterpart while occupying just 5.9GB of memory — a reduction of more than ninefold.

The new model, built on Qwen3.8 27B, employs ternary weights from the set {−1, 0, +1} combined with FP16 group-wise scaling. This technique yields 1.76 effective bits per weight and enables a 262,000-token context window for text and image inputs. The system runs under the Apache 2.0 license and supports deployment on NVIDIA GPUs via CUDA as well as Apple devices through MLX.

Compared with the earlier Bonsai 27B release, the updated version shows measurable gains in reasoning, code generation, vision tasks, and long-horizon agentic workflows. Across a suite of benchmarks covering reasoning, mathematics, coding, instruction following, and multimodal understanding, Ternary Bonsai 2 27B scores 83.9, preserving 98.2% of Qwen3.8 27B’s aggregate performance. Notably, areas that typically degrade under compression — such as coding agents, tool use, and complex reasoning — show the strongest retention.

In practical terms, the model achieves up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and consumes only 0.714 milliwatt-hours per token, making it 40% more energy-efficient than an 8-billion-parameter model running in full precision. This efficiency could allow AI assistants to perform extended coding loops, analyze private documents, or process multimodal inputs locally without frequent cloud reliance.

The development underscores a broader shift in AI deployment strategy: intelligence density — measured as benchmark performance per gigabyte of memory — is becoming a key metric. Ternary Bonsai 2 27B demonstrates that low-bit representations need not sacrifice utility, challenging assumptions about the trade-off between model size and capability.

Technical documentation, including details on compression methodology and evaluation protocols, is available in a public whitepaper. The company also offers customization services for organizations seeking to adapt the model to specific domains or hardware constraints. Contact information for collaboration is listed on its website.