PrismML has released three compact Bonsai 2 models for running AI assistants on personal devices, pitching low storage footprints while its own benchmarks show weaker scores on some tasks.
All three versions build on Qwen3.5 and are designed for reasoning, image processing and software-tool calls on computers, phones and wearables, the company said. PrismML made them available free under the Apache 2.0 license on Oct. 15 and says training used Google v5 TPUs.
A ternary scheme limits each model weight to −1, 0 or +1. PrismML lists footprints of 0.43 GB for the 2B model, 0.95 GB for 4B and 2.00 GB for 9B. The company says comparable full-precision versions would require more than nine times as much storage.
In PrismML’s detailed table, the 2B model retained 82.9% of its Qwen3.5 baseline score in knowledge and reasoning and 89% in vision. The 4B model retained 90.9% and 91.3% in those categories, while the 9B version retained 93% and 92.5%. On the GPQA Diamond benchmark, 2B scored 39.6 against 58.48 for its baseline, which PrismML gives as 67.7% retention.
Package options vary. PrismML’s 9B model card also lists a 2.42 GB language-model file in another format and a separate 0.62 GB vision projector. The announcement does not give total runtime memory requirements.
PrismML reports speeds of up to 120 tokens per second for 2B and 34 for 9B on an iPhone 18 Pro Max. It lists NVIDIA GPU support through llama.cpp and Apple-device support through MLX. Babak Hassibi, PrismML’s founder and CEO, said on-device agents must make good decisions while “working within limited memory and power budgets.”
PrismML announced its 27B Bonsai 2 model on Sept. 17, bringing the family to four sizes with this release.
