PrismML Releases Bonsai 2 Small Ternary Models, With the Smallest 9B Weight File Reduced to 2GB

On October 15, PrismML expanded its Ternary Bonsai 2 family with Qwen3.5-based 2B, 4B, and 9B models designed for on-device agents on phones, personal computers, and wearable devices. The three models combine text reasoning, visual understanding, and tool calling. Their smallest published sizes are 0.43GB, 0.95GB, and 2GB, respectively, and they are released under the Apache 2.0 license. Official announcement

The compression method represents weights using −1, 0, and +1, together with group scaling values. The 9B model card lists three storage layouts: PTQ1_0 at 2GB, PQ2_0 at 2.42GB, and Q2_0 at 2.56GB. All three preserve the same ternary weights, which the kernel reads directly without first expanding them to FP16. So the smallest file is not necessarily the fastest option, and 2GB should not be mistaken for the memory required to run the complete service. Visual input requires an additional module; the recommended Q8_0 version adds 0.62GB. 9B model card

How much capability is retained also varies by task. In the official evaluations, the 9B model scored 78.02 on BFCLv4 Mobile, compared with 82.02 for the original Qwen3.5-9B. On Tau2Bench, its score fell from 85.25 to 73.02. These figures show that performance on a tool-calling benchmark cannot be taken as evidence of equivalent reliability across an entire agent workflow. The results are from the development team and have yet to be independently reproduced. Official evaluation table

Deployment also comes with a format compatibility requirement. The model card says PTQ1_0 and PQ2_0 require PrismML’s llama.cpp fork; recent upstream versions support only Q2_0. Engineers therefore cannot assume that downloading the smallest GGUF file will make it load directly in their existing runtime. Further validation should record the runtime version, weight format, visual module, and context length, then measure end-to-end task latency and peak memory. Chinese instructions, Traditional Chinese document recognition, and multi-step tool use also need dedicated test sets. Deployment notes