PrismML announced on October 15, 2026 that it is adding three smaller Bonsai 2 releases: 2B, 4B, and 9B, built on Qwen3.5. In the PrismML announcement, the company lists footprints of 0.43GB, 0.95GB, and 2GB for those models, and says they are more than nine times smaller than comparable full-precision versions.
The new weights are available under Apache 2.0. PrismML says builders can run them through llama.cpp on NVIDIA GPUs and through MLX on Apple devices, giving teams clear local deployment paths.
For builders, that adds smaller open-weight options for private or low-latency agent workloads on phones, Macs, and other edge hardware where larger checkpoints are harder to fit.
PrismML also says it tuned the models for practical small-model behavior, including instruction following, mobile tool calling, reducing repetitive loops, and controlling verbosity. Those quality claims come from the vendor announcement, so teams will still want to test them against their own workloads.