GLiNER2.5-Decide与Jev对比:无需LLM的精准类型化决策
GLiNER2.5-Decide vs Jev: typed decisions without asking an LLM for JSON
Agent开发者的实战干货,实测证明专用分类器在路由场景下比LLM快得多且更省资源,直接给出延迟数据与代码实现细节,建议压测你的链路。
Disclosure: I maintain speech-swift, where I ported this model. Everything below is Apache-2.0.
披露:我维护着 speech-swift,在此移植了该模型。以下内容均采用 Apache-2.0 许可证。
TypeSafe's Jev and Fastino's GLiNER2.5-Decide both answer "which of these labels fits this text" with a probability per label instead of generated text. I ported Decide to MLX Swift and read both vendors' docs to see how they actually compare.
TypeSafe 的 Jev 和 Fastino 的 GLiNER2.5-Decide 均通过为每个标签提供概率来回答“哪个标签适合这段文本”,而非生成文本。我将 Decide 移植到了 MLX Swift,并阅读了两家供应商的文档以了解它们实际上的对比情况。
The headline benchmark (Decide 60.2% vs JevK5 57.6% on Fastino's fast-decisions suite) is not against TypeSafe's Jev. JevK5 is an open reproduction on a 4B-class Qwen3.5 decoder, and Fastino says so in the post. So it compares an encoder with a decoder readout, not with the commercial API.
主要的基准测试(Decide 60.2% vs JevK5 57.6%,基于 Fastino 的 fast-decisions 套件)并非针对 TypeSafe 的 Jev。JevK5 是在 4B 类 Qwen3.5 解码器上的开源复现,Fastino 在帖子中已说明这一点。因此,这是编码器与解码器读取输出的对比,而非与商业 API 的对比。
What I could measure is the local side. On an M5 Pro, INT8: 7.6 ms per routing call, 0.85 GB. All three precisions give the same labels and spans as the PyTorch original on 24 reference cases.
我能测量的是本地端的情况。在 M5 Pro 上,INT8 精度下每次路由调用耗时 7.6 毫秒,占用 0.85 GB 内存。在 24 个参考案例中,所有三种精度给出的标签和跨度均与 PyTorch 原版一致。
Where they differ: Jev takes up to 64k tokens of state and offers rubric scores and yes/no probabilities; the Swift port handles 512 tokens, single-label classification and entity spans, and runs offline.
差异点在于:Jev 支持最多 64k token 的状态,并提供评分标准和是/否概率;Swift 移植版仅处理 512 token,支持单标签分类和实体跨度,且可离线运行。
Write-up: https://soniqo.audio/blog/gliner-decide-vs-jev
详细分析:https://soniqo.audio/blog/gliner-decide-vs-jev
If you route with an LLM and structured output today, what would make you switch to a classifier like this?
如果你今天使用 LLM 进行路由和结构化输出,是什么因素会让你切换到像这样的分类器?
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力