Cloudflare推出了两个开源决策模型Clef和Clef-flash，旨在提供确定性的结构化分类结果1。这些模型在Jev Decision Index基准测试中表现领先，其中Clef在43项基准测试中的延迟性能超越所有竞争对手1。与竞争对手gpt-oss-120b模型相比，Clef在网站分类任务中的延迟降低约50%，从4.7秒降至2.2秒1。
两款模型均以Qwen作为基础模型，其中Clef使用冻结的Qwen3.8-27B，Clef-flash则采用Qwen3.5-9B1。Clef具有64k的上下文窗口，超过竞争对手Jev的32k1。与仅支持文本分类的Jev系统不同，Clef配备视觉编码器，支持图像和文本分类功能1。
这些模型已在Hugging Face上以Apache 2.0许可开源1，并通过Cloudflare的Workers AI平台托管1。Cloudflare还提供微调服务，目前由前置部署工程师团队负责，后期将推出自助式微调平台1。
Cloudflare has introduced Clef and Clef-Flash, two open-source decision models designed to deliver deterministic, structured classification outputs 1. These models are built on Qwen as the foundation, with Cloudflare freezing Qwen 3.8-27B for Clef and Qwen 3.5-9B for Clef-Flash 1. The models are released under Apache 2.0 license and available for download on Hugging Face 1.
The new models demonstrate significant performance advantages over competing systems. In the Jev Decision Index benchmark tests, Clef and Clef-Flash lead across 43 benchmark measures in latency performance 1. For practical applications, Clef achieves approximately 50% lower latency compared to the gpt-oss-120b model in website classification tasks, with response times of 2.2 seconds versus 4.7 seconds 1. Additionally, Clef features a visual encoder enabling image classification, a capability absent in its competitor Jev, which supports only text classification 1. The models also offer a larger context window of 64,000 tokens, exceeding Jev's 32,000-token limit 1.
Cloudflare is providing model fine-tuning services through a team of field deployment engineers, with plans to release a self-service fine-tuning platform in the future 1. The models are hosted through Cloudflare's Workers AI platform 1.
评论
还没有评论，欢迎留下第一条。