社区测试者称赞 Spark X2.6 4B 的工具调用表现
AI 总结
一名 Reddit 社区测试者称，Spark X2.6 4B 在 pi cli 的工具调用测试中表现良好。据该测试者介绍，模型生成了 Blender Python 脚本，通过 12 次工具调用分析复杂 Rust 代码库，找到并理解全局 agents.md 指令，并在每轮结束后列出变更文件。测试者称模型体积小、速度快，键值缓存也很小，但不支持视觉。
深度见解
该报告展示了模型在脚本编写、代码库分析、指令发现和多轮文件追踪等工具使用任务中的具体案例，呈现了社区测试者对其能力的实际观察。
https://huggingface.co/collections/XHToken/spark-x26
I tested it last night and it is amazingly good at tool calling. Tests in pi cli:
- Can write a python script for blender to create a red cube, rotate it and render it.
- It understood a rather complex rust codebase and explained details about things that work one way in one part of the app and slightly different in another part. 12 tool calls, straight to the point.
- I ran pi cli without skills and extensions and it still found the global agents.md, understood it and wrote a list of changed files after every turn. I have a global skill in ~/.agents that asks for this
I's small, fast and the kv cache is small, too. It's thinking is measured and on point. MTP should follow soon. It lacks vision.
temp: 1.0, I didn't set any other params.
What are your experiences with it?