karpo-decision-27B-act
fast decision-making model for browser agents
2 min readupdated
karpo-decision-27B-act is a specialized model by MachinePulseAI built for the System 1 (fast thinking) role within the open-source karpo-browser agent. Based on Qwen3.8-27B, it selects the most likely action on a web page in a single pass (clicks, form filling, scrolling, or task completion) based on a screenshot and element list. The model runs very fast—averaging about 931 milliseconds per decision—which significantly speeds up browser agents and reduces runtime costs compared to using heavy general-purpose language models alone.
Key features
- Web action specialization: processes text page elements and screenshots to pick the exact step.
- High speed: makes decisions in a single forward pass without text generation, averaging 0.9 seconds.
- Hybrid architecture: built on Qwen3.8-27B with an added linear readout to evaluate options.
- Cost-efficiency: substantially lowers the cost of executing tasks when paired with a more powerful System 2 model.
What you can do
- Integrate into web agents: use as a fast-response module (System 1) to automate browser actions.
- Build an LLM pipeline: combine with a large model for long-term planning (System 2) and quick site navigation.
- Automate routine workflows: set up scenarios for running tests, collecting data, and monitoring web resources.
Example
Inference example and request configuration for the model in Python:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "MachinePulseAI/karpo-decision-27B-act"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
# Пример работы с состоянием страницы и списком элементов
print(f"Модель {model_id} успешно загружена")
Read more
DetailsBenchmarks: 84% accuracy across 38 key tasks on Browserbase Benchmark v2 when paired with DeepSeek V4.1 Flash.
Parameters: 26.1B in bf16 format, plus a 255 × 5,120 readout in fp32.
Context: up to 32,768 tokens per sequence.
License: Apache-2.0.










Comments