Successors to the first-generation Bonsai 1.7B, 4B, and 8B, these models target on-device agency with focused capabilities in tool calling and vision. In a hybrid workflow demo on Mac, Bonsai 2 processed sensitive data on-device to send only 8.6% of information to the cloud, cutting cloud costs by 86%. User benchmarks show the 9B model decoding at 114 tok/s on an RTX 3060, while the 2B model reaches 500 tok/s on an M4 Max via a browser, and PrismML has launched a competition on Yukon Research to further optimize the 9B version.
Wednesday, October 21, 2026
Story
PrismML Releases Bonsai 2 Local AI Models Starting at 0.43 GB
AI6 days ago
These ternary software tools bring advanced reasoning and visual comprehension directly to smartphones and wearables to reduce reliance on external servers. PrismML has released the Bonsai 2 family of AI models based on Qwen3.5 in versions ranging from 2B to 27B parameters under the Apache 2.0 license. The 2B, 4B, and 9B versions feature weight payloads of 0.43GB, 0.95GB, and 2.0GB respectively, with the 9B variant using 40% fewer tokens for math and instruction following than its predecessors.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.