Bonsai 2 Brings Agentic AI to iPhone

Bonsai 2 makes local mobile agents feel technically plausible, although sustained battery use, runtime compatibility, and independent evaluations will matter more than headline throughput.

• –The 9B model fits into roughly 2GB, making capable multimodal inference viable on phones and other constrained devices.
• –PrismML reports 95.1% retention against the full-precision Qwen3.5 9B baseline on its mobile tool-calling benchmark.
• –Native MLX support covers Apple devices, while llama.cpp supports NVIDIA GPUs, giving developers practical deployment paths.
• –Tool calling and vision move the model beyond offline chat toward assistants that can inspect local data and perform short workflows.
• –Early community testing is enthusiastic, but the ecosystem still depends on specialized runtimes and model-loader support. [Bonsai Demo](https://github.com/PrismML-Eng/Bonsai-demo) [Community discussion](https://www.reddit.com/r/LocalLLaMA/comments/1x6w658/ternary_bonsai_2_2b_4b_and_9b_just_released_on/)