PrismML 发布了小至 0.43GB 的三值 AI 模型，供本地代理使用

免费的、采用 Apache 2.0 许可的 Bonsai 2 模型使用三元权重，并新增视觉和工具调用功能，同时附有 PrismML 报告的性能数据。

By · Published

Primary source: PR Newswire

为何重要

较小的模型文件可以使在内存和功耗受限的设备上进行本地推理更为可行。PrismML 的发布在这一主张上加入了视觉能力和工具使用功能，但其公布的主要性能数据来自公司报告，开发者应在目标硬件上对这些模型进行测试。

PrismML releases 2B, 4B and 9B ternary models for on-device agents — The free Apache 2.0 models have stated footprints of 0.43GB to 2GB; PrismML says they combine reasoning, vision and tool calling.

Babak Hassibi, a Caltech professor and co-founder and CEO of PrismML, released 2B, 4B and 9B models in the Bonsai 2 family on October 15th. Based on Qwen3.5, the models are intended to bring reasoning, visual understanding and tool use to local AI agents on computers, phones and wearables.…

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Reader comments

Conversation for this story loads after sign-in.