Back to News

MachinePulse AI open-sources Karpo Decision 31B and browser agent framework

#open-source#decision-model#browser-agent#benchmark

MachinePulse AI has open-sourced Karpo Decision 31B, a decision model for high-frequency, low-latency agent reasoning, along with its browser operation framework. In the Decision Index benchmark, the model scored 73.9 Full, ranking first among 300 models, and outperformed Gemini 3.5 Flash in a city-scanning task by being nearly 6 times faster. The browser framework, based on a System 1 + System 2 architecture, achieved an 89% average success rate on Browserbase Benchmark v2 with a per-task cost of $0.096.

Coverage timeline

  1. 量子位量子位

    反超Gemini之后,Karpo新作再开源,浏览器Agent速效双赢 衡宇 2026-11-21 19:24:29 来源: 量子位 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?! 同样一场《城市扫描仪》挑战,Karpo从36场活动中帮我挑出10个休闲娱乐的好去处(平均消费50美元之内、通勤距离仅5km)只花了2.6秒,而Gemini 3.5 Flash则足足用了14.96秒。 快了将近6倍!各种筛选条件也一个没漏,36道题准确率达100%。 这位新晋决策高手,就是MachinePulse AI刚刚开源的 Karpo Decision 31B 。 它专门为Agent提供高频、低延迟的决策能力,让每一步判断都更快、更省。 比如浏览网页下一步点哪个按钮、任务有没有完成、几十个工具待调用哪个最符合要求,它都能直接给出判断, 干起活来相当麻利。 一张RTX PRO 6000就能让它上岗,面对公开集14万多次请求, 中位响应时间仅94毫秒,眨眼的工夫都用不着。 在足足有300个模型同台厮杀的Decision Index评测总榜上, Karpo Decision 31B 还以73.9的Full分荣登榜首,单题最多可同时判断255个候选选项,校准误差降至了2.3%。 更进一步,MachinePulse还把已经应用于自家Personal Agent产品Karpo的浏览器操作框架,正式向开发者开源! 这套方案采用 System 1 + System 2 架构,包括 karpo-browser框架 、专门针对网页动作决策训练的 Karpo Decision 27B Act 模型,以及配套的网页动作决策评测集 KBB 。 在Browserbase Benchmark v2的102道核心任务上,这套方案做到了 89%的平均成功率、201秒的平均耗时 ,单任务模型成本仅需 0.096美元(约0.64元) 。 相比榜首Claude Fable 5.2 + Claude Code, 完成任务的速度快了约1.9倍,模型成本更是只有对方的十三分之一。 我们观察到,当把成功率要求放宽到80%,但真正用户可用的体验速度250秒内和成本0.15美金内,整个榜单全部91条记录无一做到,Karpo独此一份且大幅度领

Related stories

  • Google DeepMind unveils Gemini 4 Carbon, tops AI benchmark

    Google DeepMind announced Gemini 4 Carbon, its next frontier model, claiming state-of-the-art performance in cybersecurity and long-horizon tasks. The model scored 59 on the Artificial Analysis Intelligence Index v5.0, ranking first ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). It is initially available only to select cybersecurity defense companies through the Fairwind program, with pricing at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later.

    8 reports · 6 sources1013 points · 801 comments
  • Google launches Gemini Decisions API with Decision 1 model, priced at $0.10 per 1M input tokens

    Google launched the Gemini Decisions API, powered by its new Gemini Decision 1 model, which returns predefined answers with confidence scores for yes/no, choice, and rubric-scoring questions. The API is priced at $0.10 per million input tokens with free output, and supports text, images, audio, video, and PDFs. It joins Jev and OpenAI in the decision model category, with Jev remaining the cheapest at $0.042 per million tokens.

    2 reports · 2 sources
  • Anthropic launches Claude Fable 5.2 and Mythos 5.2 with coding, reasoning, vision gains

    Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, flagship models aimed at complex technical work, with claimed improvements in coding, reasoning, vision, and communication. The models are priced at $8 per million input tokens and $40 per million output tokens, a 20% cut from 5.1, with a fast mode at $16/$80. During prerelease testing, the models formalized a proof of Fermat's Last Theorem in 30 hours, solved two open black hole physics problems, and scored 64.9% on the RoboDojo Sim benchmark, up from 27.6% for Fable 5.1.

    14 reports · 7 sources916 points · 1051 comments
  • Google launches Gemini 4 Carbon to public, adds swarm feature in Antigravity

    Google DeepMind has made Gemini 4 Carbon, its self-described strongest model, available to the general public via consumer, developer, and enterprise channels, two weeks after an initial release to cybersecurity defenders. The model is priced at $2 per million input tokens and $10 per million output tokens, with a scheduled increase to $4/$20 on December 1st. Google also previewed a 'swarm' feature in its Antigravity IDE, and highlighted a research case where Carbon-powered agents formally verified Signal's encryption code.

    6 reports · 6 sources295 points · 165 comments

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe