Lightpanda Cuts Browser Memory, but Limits Concurrent AI Agent Pages
An independent developer benchmark published on Juejin compared Lightpanda with headless Chrome on the same 1-core, 2GB Ubuntu machine. Lightpanda used about 14 times less peak memory and ran 1.5 to 2 times faster on CPU-heavy pages, while network-bound pages were roughly even and the official 9x speed claim was not reproduced. The tests also found that each CDP connection permits only one browser context and one active page, and that Lightpanda does not provide graphical canvas rendering.
For teams giving AI agents lightweight web access on small machines, the benchmark indicates substantial memory savings, but existing Chrome-based concurrency designs would need separate Lightpanda processes and additional process management.
给Agent配浏览能力这件事，我之前一直用headless Chrome硬扛。手上那台1核2GB的Ubuntu小机跑着几个定时抓取任务，每次Chrome起来，free -m 掉到100MB出头的可用内存，多开两个任务就得考虑先后顺序。前几天Lightpanda刷屏的时候——就是那个用Zig从零写的无头浏览器，掘金热榜和HN都在转——README里那行"16x lighter and 9x faster than Chromium"我没直接信，因为官方基准是AWS m5.large上爬933个页面的聚合数字，跟我这种2GB小机跑零散任务的场景差得远。
正好手头环境现成，我把它跟Chrome在同一台机器上跑了一遍。结论先放这：省内存是真的，我量出来约14倍，跟官方16倍是一个量级；"9倍快"在我这个场景没复现——网络瓶颈的页面两边打平，CPU吃得重的页面Lightpanda快1.5到2倍；最值得说的是一个README里没写的坑——它的并发模型跟Chrome完全不同，代码直接迁会翻车。