PrismML releases Bonsai 2 2B, 4B and 9B, ternary models based on Qwen3.5 that fit in as little as 0.43GB for phones and laptops
Today, we’re releasing Bonsai 2 2B, 4B, and 9B. Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices: • 2B: 0.43GB • 4B: 0.95GB • 9B: 2.0GB For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait. These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables. The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware. Available today under Apache 2.0.
Context
Models this small could run useful assistants and agents locally on phones and wearables, without sending data to the cloud.
This is a first hand launch post from PrismML, and the benchmarks are its own. Its table shows the compressed models trailing the original Qwen3.5 models on knowledge, tool calling and vision while beating them on some math and instruction following tests.
Deep dive
On October 15, PrismML posted that it was releasing Bonsai 2 2B, 4B and 9B. It said the ternary models, based on Qwen3.5, bring stronger reasoning, visual understanding and tool calling to compact footprints for everyday devices, at 0.43GB, 0.95GB and 2.0GB respectively.
The company said it focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices and giving useful answers without unnecessary output. It said the 9B model uses roughly 40% fewer tokens on standard instruction following and everyday math tasks, and that the Bonsai 2 family now spans four sizes from 2B to 27B, available under Apache 2.0.
PrismML's documentation says every language model weight is ternary, one of three values, which works out to under 2 bits per weight instead of 16, and that the models take images, carry a 262K token context and call tools natively. Its blog says the models run on Nvidia GPUs through llama.cpp and on Macs, iPhones and iPads through MLX. The docs warn that older llama.cpp builds can load the files but output gibberish.
In PrismML's benchmark table, Bonsai 9B scores 74.46 on knowledge and reasoning against 80.10 for Qwen3.5 9B, but 92.06 on math against 93.71.
What to expect next
Watch for independent tests of how much quality the ternary models lose compared with the original Qwen3.5 models in real assistant use.
Also watch whether phone makers or app developers ship Bonsai 2 in local features, and whether upstream llama.cpp and MLX fully support PrismML's formats so the models are easy to run. Speed tests on ordinary phones will be the key practical check.