Kev：基于Qwen3.5的轻量级决策模型系列发布

Kev是一个新开源项目，提供了基于Qwen3.5构建的轻量级决策模型系列1。该系列包含0.8B、4B和9B三个版本1，可处理是/否、多选和评分等多种决策任务1。项目采用Apache-2.0开源许可证1，为用户提供预训练权重、本地部署支持和Web演示1。

在性能表现上，Kev-9B在新数据集开发集上的精度相比Kev-8B提高了7.3个百分点（95%置信区间为+2.8到+11.7）1，Brier分数降低了0.081。在硬件方面，4B和9B版本可在32GB Mac上以bf16精度运行1，项目同时支持Apple Silicon和CUDA平台1。模型支持通过LoRA适配器进行微调，用户可在特定领域对模型进行定制训练，同时保留预训练知识1。

Jared Palmer has released Kev, an open-source family of small decision models constructed on top of Qwen3.5 for handling classification and scoring tasks 1. The project includes three model sizes—0.8B, 4B, and 9B parameters—each capable of processing binary yes/no questions, multiple-choice selections, and scoring assignments 1.

The models are available under the Apache 2.0 license with open-source code, pretrained weights, and local deployment support 1. The 4B and 9B versions can run on 32GB Mac machines using bfloat16 precision 1. On a newly developed evaluation set, Kev-9B achieved accuracy 7.3 percentage points higher than Kev-8B (95% confidence interval: +2.8 to +11.7), and posted a Brier score 0.08 lower than its predecessor 1. The project supports fine-tuning on domain-specific data through LoRA adapters with rank-16 configuration while preserving pretrained knowledge via the --init_from parameter 1. Both Apple Silicon and CUDA acceleration are supported for inference 1.