Ternary Bonsai 2 is PrismML's family of ternary-weight models built for on-device deployment. On October 15, 2026, PrismML announced three additions to the line, a 2B, a 4B and a 9B, all based on Qwen3.5, taking the family to four sizes spanning 2B to 27B. All three are available now as free downloads under the Apache 2.0 license.
Ternary weights shrink the size to a tenth
A conventional model's full-precision weights use 16 bits per parameter. Bonsai 2 instead stores each weight as one of three values, -1, 0 or +1. The result is a steep drop in footprint: the 2B model is 0.43GB, the 4B is 0.95GB and the 9B just 2.00GB, each less than a ninth of its full-precision counterpart. This continues the route PrismML showed earlier this year running a 1-bit vision model on smart glasses, a demo this site covered when the 1-bit model reached Snapdragon glasses. Small size is not only about saving storage: smaller weights mean less memory-bandwidth pressure, which is what gives ordinary CPUs and edge GPUs a chance to keep latency down.
How fast it runs depends on the device
According to PrismML's figures, on hardware such as an RTX 4090 or a MacBook M5 Pro, Bonsai 2 reaches up to about 3 times the energy efficiency and 5 times the throughput of the matching full-precision models, with the 9B at up to 124 tokens per second and the 2B at up to 276. On an iPhone 18 Pro Max, the 9B runs at 34 tokens per second and the 2B at 120. On capability, the company says that across 19 benchmarks the new models match or exceed their full-precision counterparts on coding, math and instruction following, and retain more than 90 percent of full-precision Qwen3.5 scores on vision, agentic and tool-use, and knowledge and reasoning tasks. These are vendor-measured numbers, so real workloads still need testing on your own data.
Who should download and try them
For teams building offline assistants on phones, glasses or laptops, these three are a rare option: not small in parameters, yet tiny in size, able to complete short task chains through tool calling without a network connection or sending data to the cloud. If your workloads already run on GPU servers and you are chasing maximum quality instead, the space ternary compression saves matters little, and a full-precision large model remains the safer pick. Competition among on-device agents is shifting from whether a model can run at all to how much capability fits inside such a small footprint, and Bonsai 2 just pushed that size table one notch lower.
