PrismML releases Bonsai 2 27B, compressing Qwen3.8 to a 5.9GB footprint

PrismML released Bonsai 2 27B on September 17, 2026. [2][3] The startup, led by CEO Babak Hassibi and backed by a $22.25 million seed round from investors including Khosla Ventures, compressed Alibaba's open source Qwen3.8 27B model down to 5.9 gigabytes. [2] This represents a 9x to 10x reduction in memory compared to the original model. [2] The company reports that the compressed model retains 98.2% of the full-precision version's aggregate benchmark scores. [1][3] PrismML states the model can achieve 143 tokens per second on an RTX 5090 graphics card, though running it currently requires a proprietary fork of the llama.cpp software. [3] The AI lab is betting that reasoning models can be made small enough to fit on standard personal computers and smartphones. [2]

The release of Bonsai 2 27B illustrates a continuing push to run capable reasoning models locally on consumer hardware. By matching Alibaba Qwen3.8's performance closely at a fraction of the size, PrismML provides a direct alternative to cloud-reliant inference. This development compares favorably to previous compression efforts by minimizing benchmark degradation, offering a viable path for on-device processing. What the sources don't address: whether independent real-world testing will consistently validate the company's high throughput claims across diverse consumer hardware setups.