PrismML has released three compact, open-weight Bonsai 2 models designed to run locally, including a 9B model with a stated footprint of about 2GB. The company says they combine reasoning, vision and tool use with lower memory and energy demands, making local assistants more practical on everyday devices.
PrismML Watch analysis
What happened
The new ternary models come in 2B, 4B and 9B sizes, joining PrismML’s existing 27B model. Their weights use three values, {-1, 0, +1}; PrismML gives the respective model footprints as 0.43GB, 0.95GB and 2GB. The weights are available under the Apache 2.0 licence, and the models run on NVIDIA GPUs through llama.cpp and on Apple devices through MLX.
Our top picks
- Small footprints
The 2B, 4B and 9B models are listed at 0.43GB, 0.95GB and 2GB, respectively, lowering the memory barrier for local use. - Tool use on mobile
PrismML highlights tool calling, including mobile tool use, alongside instruction following and visual understanding. - A 9B model for phones
The company says Bonsai 2 9B reaches up to 34 tokens per second on an iPhone 18 Pro Max. - Higher throughput on selected hardware
PrismML reports up to five times the throughput and three times the energy efficiency of the corresponding full-precision models on RTX 4090 and M5 Pro hardware. - Open weights and broad device support
Apache 2.0 licensing, llama.cpp support for NVIDIA GPUs and MLX support for Apple devices give developers several routes to try the models locally.
Why it matters
Running an assistant on a phone or computer can avoid sending every task to a remote service, but local models have to make sensible trade-offs between capability, speed and limited memory or power. PrismML’s release targets that constraint directly, with several sizes and support for tool-driven and visual tasks.
The performance figures and benchmark comparisons come from PrismML, not an independent evaluation. The results are promising, but developers will want to compare them on their own workloads before treating headline speed or benchmark retention as a guarantee.
Our read
This is more than a smaller-model announcement: the combination of open weights, device support and reported mobile performance gives developers something concrete to try. The 9B model is the obvious test case for richer local assistants; the smaller sizes may suit narrower jobs where memory is the stricter limit. If you build locally, the useful next step is to test the model on your own device and tasks, not to let a benchmark table pick your assistant for you.
What to watch
- How the models perform on independent tests and ordinary user tasks.
- Whether tool use and visual understanding hold up on mobile hardware.
- How model quality, speed and battery use compare across the three sizes.
Discussion spark: For a local assistant on a phone, would you choose a smaller model that responds quickly and uses less power, or a larger one with more capability even if it needs more memory and battery?
Sources and evidence
- PrismML – Introducing Ternary Bonsai 2 2B, 4B, and 9B: Bringing Local Agents to Everyday Devices (15 October 2026, 00:00 UTC)
not affiliated with or endorsed by PrismML