MachinePulse AI open-sources Karpo decision models and browser-agent tools
MachinePulse AI open-sourced six Karpo Decision models, the Karpo Decision 27B Act web-action model, the TypeScript karpo-browser framework, and the KBB browser-action benchmark. The browser system uses a general-purpose model for planning and the dedicated action model for web interactions, while karpo-browser handles page reading, execution, page changes, and errors. The decision models use a Readout Head to return candidate probabilities in one forward pass without token-by-token generation. QbitAI reported that the team’s tests found an 89% average success rate across 102 Browserbase Benchmark v2 tasks, with an average duration of 201 seconds and model cost of $0.096 per task.
The release provides an open toolchain spanning model weights, browser execution software, and benchmark data. Its split between general-purpose planning and dedicated web-action decisions is designed to avoid generating answers token by token for frequent agent choices, giving developers a way to reduce unnecessary generation work.
反超Gemini之后,Karpo新作再开源,浏览器Agent速效双赢 – 量子位
< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">
反超Gemini之后,Karpo新作再开源,浏览器Agent速效双赢

文婷 发自 凹非寺 量子位 | 公众号 QbitAI
2026年了,居然还有年轻团队专门训练决策模型,而且一出手,就把Gemini甩在了身后?!
#MachinePulse AI#Karpo Decision 31B#Karpo Decision 27B Act#karpo-browser#Karpo BrowserBench#DeepSeek V4.1 Flash#Gemini 3.5 Flash#Claude Fable 5.2#Claude Code#Codex