Mano-CUA's ten-point win over Claude shrinks to 0.8 points against Gemini

The 10.4-point spread the post leans on is 41.7 minus 31.3 ^[8][1][3]. Take the same three numbers from the other end. Gemini 2.5 Pro Computer Use scored 40.9, which puts it 9.6 points above Claude 4.5 Computer Use and 0.8 points below Mano-CUA 1.1 ^[2][10][9]. Nine point six of those 10.4 points separate Claude from the other two entries ^[10][8].

"General-purpose LLMs are bad at browser automation," the post's author wrote ^[7]. The post sets up the comparison as specialized against general-purpose, with Mano-CUA on one side and Gemini and Claude on the other ^[21]. On that framing, the premium a purpose-built model buys over the better general entry is 0.8 NavEval points ^[9].

Mano-P plans from screenshots: click targets are identified from pixels, with no DOM parsing, no API hooks and no accessibility tree ^[11]. The post says most general models parse the DOM or lean on accessibility APIs, and that this breaks on dynamic rendering, custom components and heavy JavaScript ^[12]. It also describes a think-act-verify loop, so a dropdown click that opens nothing is retried, and a field that rejects input has its error message read back off the screenshot ^[13].

For 41.7 to transfer, your pages have to fail the way the benchmark's pages fail. WebRetriever Protocol I draws on real forms from live sites, including captchas, multi-page flows and validation errors that send you back to fix something ^[4]. The failure the post attributes to general models is losing track of position in a form after four or five steps ^[14]. If your target is an internal app with stable test IDs and short forms, that is not a cost you are currently paying. The post does not say how many tasks NavEval covers or how many runs each model got.

Install is a brew tap, then check, install-sdk, install-model and run ^[16]. The 4B quantized build wants a Mac with Apple silicon and 32GB of RAM, and decodes at about 80 tokens per second on M5 Pro hardware ^[15][16]. At that rate, a step whose plan is 100 output tokens costs 1.25 seconds of decode before the page has even reloaded ^[22].

The model and the benchmark setup ship from the same repository, github.com/Mininglamp-AI/Mano-P ^[5], and the scores were published on Mininglamp's dev.to account ^[6]. The losing configuration was therefore set up by the party publishing the win. I would want to run that leaderboard myself before moving a pipeline onto it. The post describes the training story but does not measure it: supervised fine-tuning on GUI interaction data, offline reinforcement learning on logged trajectories, then online RL against live interfaces ^[18].

The local-only guarantee is one command-line flag. With --local, inference stays on the machine and no screenshots leave it ^[15][16]. Without it, inference goes through Mininglamp's cloud endpoint, which the post says is faster, and the screenshots go to their server ^[17].