Strands发布开源决策模型Strands Decider 2B

Strands公司推出了一款名为Strands Decider 2B的开源决策模型，专门针对快速决策和代理应用场景设计1。该模型拥有20亿参数，可在本地CPU或GPU上运行1。在决策延迟性能方面，中位决策延迟在Nvidia RTX3090上约为115毫秒，在MacBook M3上约为153毫秒1。

模型在准确性和校准度表现竞争力强1。在JevBench公开基准集上，该模型在33个2B级模型中排名第3，排除超过2B参数的模型后排名第11。在简单任务上可达到100%的准确率1。技术上，该模型基于Qwen3.5-2B预训练模型，通过rank-16 LoRA适配器微调得到，已迭代至v19版本1。

Strands已将所有代码、权重和训练数据在GitHub和Hugging Face上开源1。

Strands has released Strands Decider 2B, an open-source artificial intelligence model designed for rapid decision-making and agent applications 1. The model contains 2 billion parameters and runs efficiently on local CPU or GPU hardware 1.

The system demonstrates strong performance across multiple metrics 1. Median decision latency reaches approximately 115 milliseconds on an Nvidia RTX 3090 and about 153 milliseconds on a MacBook M3 1. On the public JevBench benchmark, the model ranks third among 33 models in its parameter class, and first when excluding models exceeding the 2-billion-parameter threshold 1. The model achieves competitive accuracy and calibration scores, with 100% accuracy on certain simpler tasks 1.

Built upon the Qwen 3.5-2B pretrained foundation, Strands Decider 2B was refined using rank-16 LoRA adapter fine-tuning and has progressed to version 19 1. All development resources—including code, model weights, training data, and scripts—are publicly available on GitHub and Hugging Face 1.