PrismML releases compact LLM with minimal performance loss
AI startup PrismML has launched Bonsai 2 27B, a highly compressed version of Alibaba's Qwen3.8 model that fits on PCs and smartphones while retaining nearly all its original capabilities. The new model reduces memory usage by up to 10 times compared to the source model, achieving 98% benchmark parity — an improvement from the previous Bonsai release in March which matched 95%. PrismML's approach uses ternary weights (values of +1, -1, or 0) instead of traditional 16-bit representations. The company plans to apply this compression technique to even larger models soon. Founded by Caltech researchers and led by CEO Babak Hassibi, PrismML has raised $22.25 million in seed funding from Khosla Ventures, Cerberus Capital, and others. While not the only player in LLM compression (Multiverse Computing is another), PrismML claims its models retain more performance than competitors'. The startup's smaller Bonsai variants have already been downloaded over 13.6 million times combined. The technology could enable advanced AI to run locally on consumer devices, improving privacy and reducing reliance on cloud services.
Topics: Bonsai model release LLM compression PrismML startup funding