PrismML expanded its Bonsai 2 family on October 15, 2026 with three new ternary models — 2B, 4B and 9B — based on Qwen3.5. Joining the 27B flagship released in September, the family now spans four sizes from 2B to 27B parameters, all aimed at one problem: putting genuinely capable AI agents inside the memory and power budgets of everyday devices.

The trick is a ternary weight representation: instead of full-precision numbers, every weight is stored as one of three values — −1, 0 or +1. That shrinks the footprints to 0.43GB for the 2B model, 0.95GB for the 4B, and 2.00GB for the 9B — more than 9× smaller than their full-precision counterparts, according to PrismML. A 9-billion-parameter model you can store in 2GB of phone memory is a genuinely new option for mobile developers.

Capability-wise, these aren't bare chatbots. PrismML says the new models combine reasoning, visual understanding, and tool calling — the trio needed for local agentic workflows: interpreting a request, looking at text or visual context, and chaining short sequences of actions through application tools. The target hardware is explicit: PCs, phones, smart glasses and wearables. And they're free under the Apache 2.0 license, available since October 15 — you can download and ship them commercially without a licensing conversation.

Performance claims need the usual asterisk: they're company-reported. Across 19 benchmarks, PrismML says the models match or exceed their full-precision counterparts on coding, math and instruction following, and retain over 90% of baseline scores on vision, agentic work, tool calling and knowledge/reasoning. Independent read-throughs of PrismML's detailed table show the smaller sizes lose more ground — the 2B model reportedly retains about 83% on knowledge and reasoning and 89% on vision versus its Qwen3.5 baseline. Real, but not lossless — benchmark before depending on it.

Speeds are the other half of the pitch. PrismML reports up to 276 tokens/second for the 2B and 124 for the 9B on an RTX 4090 or MacBook M5 Pro, and — more interesting for the mobile story — up to 120 toks/s (2B) and 34 toks/s (9B) on an iPhone 18 Pro Max, via llama.cpp and MLX support. At 34 tokens a second, the 9B is usable for real on-device assistant work.

Taken with NVIDIA's RTX Spark laptops landing the same week, the direction is clear: inference is moving off the cloud. For developers prototyping local agents on cheap hardware, Bonsai 2 is currently the strongest free option to test first.

Sources