PrismML open-sources Ternary Bonsai 2 2B/4B/9B under Apache 2.0: Qwen3.5-based on-device models with vision and tool calling in 0.43–2.0 GB

Why it matters

【Key points】

On Oct 15 PrismML, founded by a Caltech team, expanded its Ternary Bonsai 2 family with 2B, 4B and 9B models, free to download under Apache 2.0, with weights on Hugging Face. The new models are based on Qwen3.5 and store weights as one of three values (−1, 0, +1), giving footprints of 0.43 GB, 0.95 GB and 2.00 GB, more than 9x smaller than full precision. The family now has four sizes, from 2B to 27B.

PrismML's figures: across 19 benchmarks the models match or beat full-precision Qwen3.5 on coding, math and instruction following, and keep over 90% of the scores on vision, agentic and tool-calling, and knowledge and reasoning tasks. On an RTX 4090 and an M5 Pro MacBook they deliver up to ~3x the energy efficiency and ~5x the throughput of full precision, with 9B reaching up to 124 tok/s and 2B up to 276 tok/s. On an iPhone 18 Pro Max, 9B runs at ~34 tok/s and 2B at ~120 tok/s. The models were trained on Google v5 TPUs. Local apps such as Atomic Chat and Darkbloom added them on day one.

【Verifiable】

PrismML announcement: https://prismml.com/news/prismml-expands-bonsai-2-family-with-2b-4b-and-9b-models-for-on-device-ai

Hugging Face collection: https://huggingface.co/collections/prism-ml/bonsai-2

【Our read】

1. Ternary quantization squeezes a 9B model with vision and tool calling into 2 GB, which puts a small agent that can actually do work inside the memory budget of phones and glasses. That is closer to real on-device deployment than chasing parameter counts.

2. All results are vendor-reported and compared against its own Qwen3.5 base, so they need community replication. "Over 90% retained" also means vision and tool calling still lose something.

3. It is another case of a US company compressing a Chinese open base model (Qwen), a sign that the downstream ecosystem for open weights is expanding onto devices.