PrismML compresses 9-billion-parameter model weights to 2GB: generating 34 tokens per second locally on iPhone

Beating AI News Flash: U.S. AI company PrismML has added three new models to the Bonsai 2 family — 2B, 4B, and 9B — all based on Alibaba's Qwen3.5. The team constrained the main weights of the language models to three values: -1, 0, and +1, shrinking the model files to 0.43GB, 0.95GB, and 2GB respectively — roughly one-ninth the size of the original 16-bit weights. The new models still support reasoning, image understanding, and tool calling, and are primarily aimed at local Agents on devices such as phones and personal computers.

PrismML claims that on an iPhone 18 Pro Max, the 9B model generates up to 34 tokens per second, while the 2B reaches 120.

In official tests, the 9B's coding and instruction-following capabilities nearly match the original Qwen3.5 at the same scale, but the gap is pronounced on complex tasks. On Tau2Bench, which tests multi-turn dialogue and tool operation, Bonsai 2 9B scored 73.02 versus the original's 85.25; on complex document parsing, it scored 73.8 versus the original's 88.73.

All three models have been released as open weights under the Apache 2.0 license.