🌟 OpenJarvis: A Universal Framework for Creating AI Agents
Researchers from Stanford SAIL conducted efficiency measurements of local language models, assessing how well they convert electricity into real computational tasks. They called this metric “intelligence per watt” — defining the level of energy efficiency.
During the experiment, over one million queries were processed using more than 20 different models running on 8 types of accelerators. The results showed that from 2023 to 2025, the efficiency of local inference increased by a factor of 5 or more, and modern small models can already successfully handle approximately 89% of standard chat and logical queries. Hardware and algorithms are ready for use; the remaining question is software.
And now, OpenJarvis has appeared — an open platform that transforms these insights into a complete infrastructure for personal AI agents running directly on the user’s device.
The creators compare it to PyTorch: their goal is to make OpenJarvis the same for local AI as PyTorch became for deep learning — the industry standard upon which all other solutions are built.
The framework is divided into five key components:
🟢 Intelligence — a layer of language models with a unified catalog, where there’s no need to manually track updates or manage memory.
🟢 Engine — inference engine: Ollama, vLLM, SGLang, llama.cpp, Apple Foundation Models, and others. OpenJarvis automatically detects hardware specifications and recommends optimal settings.
🟢 Agents — behavior layer: includes roles for orchestrator and scenario executor with limited context and memory on the device.
🟢 Tools & Memory — integrations via MCP and Google A2A: semantic indexing of local documents, connection to messengers like iMessage or Telegram, and other capabilities.
🟢 Learning — adaptation training mechanism: local data is transformed into training samples using methods like SFT, LoRA, and GRPO. All this work is automated within the system.
A distinctive feature of the project is its approach to energy efficiency. OpenJarvis profiles power consumption on devices with NVIDIA, AMD, and Apple Silicon hardware at intervals of just 50 milliseconds.
Deployment is available via command line interface (CLI), web browser, or desktop application for macOS, Linux, and Windows.
⚠️ To enable full functionality (e.g., security or agent management), Rust must be installed.
Additionally, the team organized a leaderboard contest focused on saving resources, energy, and computing power. Anyone can participate — the winner will receive a Mac Mini as a prize.
The project license is Apache 2.0.
Additional materials:
– project article
– documentation
– community on Discord
– GitHub repository
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby