PrismML has released Ternary Bonsai 2 27B, a compressed language model that achieves a 9x reduction in size compared to its full-precision counterpart while maintaining near-lossless performance. Based on Qwen 3.8 27B, the model uses ternary weights ({−1, 0, +1}) with FP16 group-wise scaling to achieve 1.76 effective bits per weight, resulting in a total footprint of 5.9GB.
According to PrismML, Ternary Bonsai 2 27B retains 98.2% of its full-precision baseline’s aggregate benchmark performance across reasoning, math, coding, instruction following, vision, and agentic tool use. The model scores 83.9 on this benchmark suite. The company emphasizes that the capability retention is particularly strong in areas sensitive to degradation, including coding agents, tool-use systems, multimodal workflows, and long-horizon tasks.
The model supports a 262K-token context window and multimodal text-and-image input. It is released under the Apache 2.0 license and runs on NVIDIA GPUs via CUDA and on Apple devices via MLX through custom low-bit kernels.
In terms of performance, Ternary Bonsai 2 27B reaches up to 143 tokens per second on NVIDIA GeForce RTX 5090 and 46.8 tokens per second on M5 Max. On an RTX 4090, the model consumes 0.714 mWh per token, making it 40% more energy-efficient than an 8B model running in full precision.
Compared to the original Bonsai 27B released two months earlier, the new version improves capability retention from 95% to over 98%, uses a stronger base model, and delivers better reasoning, coding, vision, and agentic performance. PrismML suggests the compression enables new applications including coding-agent loops, computer-use workflows, private document analysis, and hybrid systems where local models handle sensitive or high-frequency tasks.
The release reflects a broader shift in how AI systems are deployed, according to PrismML. The company argues that the focus should increasingly be on “how much useful intelligence can be delivered within a given memory, compute, and power budget” rather than model capability alone. This approach has implications for deployment across personal devices, workstations, and datacenters.
Key facts
- Ternary Bonsai 2 27B compresses to 5.9GB, 9x smaller than full-precision baseline
- Model retains 98.2% of full-precision performance on aggregate benchmarks
- Achieves 143 tokens/second on RTX 5090, 46.8 tokens/second on M5 Max
- 40% more energy-efficient than full-precision 8B models on RTX 4090
- Supports 262K-token context window and multimodal text-and-image input
- Uses ternary weights with FP16 group-wise scaling for 1.76 effective bits per weight
- Released under Apache 2.0 license