ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial2 min readDisconfirmAI Editorial

MachinePulse AI Open-Sources Karpo Decision 31B and Browser Agent Framework

MachinePulse AI released the open-source Karpo Decision 31B model and karpo-browser framework, claiming significant speed and cost advantages for agent decision-making over general LLMs.

2026-11-21

MachinePulse AI has open-sourced Karpo Decision 31B, a specialized decision model designed to provide high-frequency, low-latency judgment capabilities for AI agents. The company also released the karpo-browser framework, which employs a System 1 + System 2 architecture to separate complex reasoning from routine web actions.

According to MachinePulse's self-reported benchmarks, Karpo Decision 31B achieved a median response time of 94 milliseconds on an RTX PRO 6000 GPU across 140,000 requests. In a specific 'City Scanner' challenge involving 36 activities, the model completed the task in 2.6 seconds with 100% accuracy, compared to 14.96 seconds for Gemini 3.5 Flash. On the Decision Index leaderboard, Karpo scored 73.9, leading among 300 models.

The karpo-browser solution integrates the Karpo Decision 27B Act model for web navigation decisions. MachinePulse reports that this setup achieved an 89% average success rate on the Browserbase Benchmark v2 (102 tasks) with an average duration of 201 seconds and a model cost of $0.096 per task. This is claimed to be approximately 1.9 times faster than Claude Fable 5.2 + Claude Code, at one-thirteenth of the cost.

Technically, Karpo utilizes a Readout Head mechanism to calculate probabilities for multiple options in a single forward pass without generating text tokens, reducing output latency. It also employs Branch Attention to share context across multiple questions, allowing batched decision-making. The team introduced Recursive Self-Improvement (RSI) training to address generalization issues, reporting a 1.6 percentage point increase in accuracy on unfamiliar domains after model fusion.

The release includes six model sizes (0.8B to 31B), the TypeScript-based karpo-browser framework, and the KBB evaluation dataset covering 584 real websites. These tools are available via Hugging Face and GitHub.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Sources

  • qbitai.com量子位↗
Continue researching →