Lightpanda's 9x Speed Claim Is Vendor-Run, and Its Own Agent Test Shows No Engine Gain
Software / explainer
Lightpanda's 9x Speed Claim Is Vendor-Run, and Its Own Agent Test Shows No Engine Gain
The Zig-built headless browser has 38.1k GitHub stars. Its docs and README disagree on the test machine, and the AssistantBench result credits tooling, not the browser.
Lightpanda, an open-source headless browser written in Zig, says it is 9x faster and 16x lighter than Chromium. Every figure behind that claim comes from Lightpanda itself, and an independent review published in August says it did not reproduce them.
The project's GitHub repository showed 38.1k stars, 1.9k forks and 11,413 commits when fetched on October 30. The licence is AGPL-3.0. Stars are not adoption, and the benchmark pages are worth reading more carefully than the star count.
What the benchmark measured
The numbers come from Lightpanda's benchmark documentation. The first test crawled 933 URLs of a demo product catalog with a Go program using chromedp over the Chrome DevTools Protocol. Chrome ran one browser with several tabs. Lightpanda ran several isolated processes, because it cannot open multiple tabs in one process.
The docs say the machine was a fresh Ubuntu AWS m5.xlarge and the Chrome version was 143.0.7499.169.
- Lightpanda4.81 seconds
- Chrome46.7 seconds
Source: Lightpanda benchmark documentation (vendor-run), accessed 2026-10-30
Peak memory in the same run was 123.0MB for Lightpanda and 2.0GB for Chrome. A second test ran a Puppeteer script 100 times against a homemade e-commerce page with two XHR requests, served locally, so network latency is excluded. Lightpanda averaged 16ms a run and peaked at 21.2M of memory; Chrome 143.0.7499.109 averaged 185ms and peaked at 402.1M.
Where the pages disagree
The README on GitHub describes the same crawl differently. It says the test requested 933 real web pages from an AWS EC2 m5.large, and rounds the result to 5s against 46s and 123MB against 2GB. The documentation says the catalog was a demo and the instance an m5.xlarge, which has twice the memory and vCPUs of an m5.large. Both cannot describe the same run, and neither page dates it.
The docs admit that the crawl compares tabs in Chrome against processes in Lightpanda, and that Chrome's duration plateaus after 5 tabs. At one tab, Chrome took 1 minute 22.83 seconds and 1.3G. The page is published by Lightpanda and says its results are not independently verified.
The outside review
Wavect, an AI engineering consultancy, published a review by Kevin Riedl on August 1, updated September 2. It calls the benchmark "designed and run by Lightpanda" and says "we did not independently reproduce it". It lists Lightpanda as beta software with no graphical rendering engine, calls its PNG export a text-layout approximation and warns that connecting over CDP proves reachability, not compatibility.
Wavect recommends a pilot for high-concurrency extraction and research agents that read markdown, and says screenshots, PDFs, visual regression and browser extensions should stay on Chrome. It adds that the AGPL licence carries network-use obligations, with proprietary licensing available from the vendor.
The agent result
The most useful table on the benchmark page is the third. Lightpanda ran AssistantBench's 33-task validation split with Claude Sonnet 4.6, a 1,800-second timeout per task, and graded final answers against known correct ones.
| Setup | Strict accuracy |
|---|---|
| lightpanda agent | 69.7% |
| Lightpanda MCP | 66.7% |
| agent-browser + Chromium | 57.6% |
| agent-browser + Lightpanda | 57.6% |
| browser-use + Chromium | 39.4% |
Swapping Chromium for Lightpanda under agent-browser changed nothing, 57.6% both ways. The page credits the 9-point lift from Lightpanda's own MCP tools to the tool surface, not the engine. That is the page's own reading of its own data, and a fair one: on this test the browser engine did not buy accuracy.
The text-only output also has a cost. On GAIA, the page reports 81.1% against 84.9% under agent-browser, a few points lost to content a rendered page would show.
Limits to know before a pilot
The README lists no native Windows binary, no native Android build and glibc-linked Linux binaries that fail on musl distributions such as Alpine. It says telemetry is on by default and can be switched off with LIGHTPANDA_DISABLE_TELEMETRY=true. Building needs Zig 0.17.0, and the browser depends on V8, libcurl and html5ever. The CDP server works with Puppeteer, and the README gives no compatibility matrix.
The README names no individual maintainers. Contributions require a CLA, and the Releases section was empty when fetched, with a rolling nightly build the only release referenced. A team weighing it should run its own corpus, pin a build and keep Chrome as fallback, which is Wavect's advice too. For other local-first developer tools, see an emulator of the Databricks API and a browser SQLite editor.
Sources
More in Software
- 01Python 3.15 Ships a Built-In sentinel After Five Years of PEP 661The new type replaces the object() idiom with a one-line constructor that survives copying, pickling and type checking, but a Python 3.15 floor is the price of using it.
- 02Lightpanda 1.0 Passes 80% of Chrome's Web Tests, by Its Own CountThe Zig-built headless browser left beta on Oct. 2, and its own benchmarks put it ahead of two Rust rivals on a 940-page crawl, with no hardware listed.
- 03Claude-Mem Hits 107,900 Stars, and Its Own Pages Disagree on LicenceThe Claude Code memory plugin has gained about 35,500 stars since May, but a May write-up says AGPL-3.0, the repository now says Apache-2.0, and the README describes an optional cloud sync that earlier coverage did not.
- 04Jepsen Found 96 Issues in ParticleDB 1.1.4 and Tells Users to Halt DeploymentKyle Kingsbury’s 31 October report, funded by vendor Orkes, finds lost writes and ignored constraints in healthy single-node runs.