NEX is a next-generation agentic model system and open-source ecosystem jointly released by Shanghai Innovation Institute along with numerous startup partners including Shanghai Miracle Plus Intelligence Summit, Moosi Intelligence, Kuafu Technology, and others. This project aims to build a sustainable, iterative, agentic closed-loop open-source ecosystem, empowering industrial upgrades through technological breakthroughs, and truly driving the arrival of the AI agency era.
As a full-stack AI Agent solution integrating models, data, Agent development frameworks, and infrastructure code, NEX is committed to significantly lowering the development and deployment barriers for AI Agents, providing academia researchers and industry entrepreneurs with high-performance, high-stability, low-cost "out-of-the-box" agentic systems, helping developers rapidly implement AI agency across various application scenarios.
- Full-Size Model Matrix: Covers the full spectrum from 8B to 671B models, flexibly meeting diverse needs from lightweight embedded scenarios to high-performance complex tasks
- Strong Agentic Performance: Achieves industry-leading levels in complex tasks such as general capabilities, agentic coding, web search, and tool calling, with efficient problem-solving capabilities
- High Practical Value: Demonstrates excellent performance in real productivity scenarios such as mini-program development, web writing, slide production, and role-playing, enhancing work efficiency
- End-to-End Autonomy: Developers can complete the full process of AI Agent development and product construction based on NEX, from data construction, model training, Agent development to deployment, achieving autonomy and reducing development costs
- Full-Chain Open Source: One-click access to data synthesis pipeline, high-quality training datasets, full-size agentic models Nex-N1, Agent development framework NexAU, MoE inference service framework EaaS, reinforcement learning training service framework NexRL, with a complete open-source ecosystem
- Scenario-Driven Innovation: Deeply integrated with real scenario needs in academic research and industrial development, focusing on optimizing the model's agentic performance for practical applications
- Ecosystem Synergy: Shanghai Innovation Institute unites research institutions and startup ecosystem partners to build an open collaboration platform, jointly exploring the boundaries of AI agency technology and co-creating industrial value loops
Based on the full-chain solution for agentic models, we open-source 4 agentic models:
| Model Name | Base Model | Link |
|---|---|---|
| DeepSeek-V3.1-Nex-N1 | DeepSeek-V3.1-Base | 🤗 HF |
| Qwen3-32B-Nex-N1 | Qwen3-32B | 🤗 HF |
| Qwen3-30B-A3B-Nex-N1 | Qwen3-30B-A3B-Base | 🤗 HF |
| internlm3-8B-Nex-N1 | internlm3-8b-instruct | 🤗 HF |
We selected 6 representative benchmark tests to comprehensively evaluate the performance of Nex-N1 models in Agent-related tasks from both general and professional capability dimensions. Results show that the Nex-N1 series models perform excellently in core tasks such as agentic coding, tool usage, and backend development.
NexAU (Nex Agent Universe) is designed with the core principles of "low barrier, high efficiency, and flexible customization," comprehensively covering the basic capabilities needed for agent development, including:
- Dynamic tool injection
- Skills & Subagents
- Hooks & MCP
- State management
- Interleaved Thinking
- Memory compression and recall
- Complete observability
This framework transforms agent construction from "complex coding" to "rapid assembly," helping newcomers get started quickly while multiplying the efficiency of experienced developers.
Example Agents based on NexAU:
- Meta Agent (Agent4Agent): GitHub
- Deep Research Agent: GitHub
- Web Writing & Academic Poster Generation Agent: GitHub
We provide an end-to-end Agentic data synthesis pipeline, utilizing the flexible and efficient NexAU agent framework and NexA4A's powerful and diverse agent construction capabilities. The pipeline comprehensively covers core areas such as agent basic capabilities, tool calling, MCP calling, and Agentic Coding.
Resources:
NexRL is a reinforcement learning training framework designed with extreme loose coupling. We have decoupled Rollout and Training through service-oriented design, seamlessly integrating with existing training and inference ecosystems through standardized APIs. Components within the framework are deeply modularized, supporting flexible replacement, and can quickly launch reinforcement learning training adapted to multi-form Agents without modifying Agent sampling code.
GitHub Repository - NexVenusCL
EaaS is a high-efficiency inference framework designed specifically for MoE models. The system achieves fine-grained elastic resource scaling and high fault tolerance by decoupling expert modules into independent, stateless services. Leveraging high-performance GPU point-to-point communication technology, the EaaS system balances high throughput with low overhead, saving 37.5% of computing resources while achieving current SOTA inference performance, and controlling throughput degradation to within 2% in hardware failure scenarios.
The open-sourced NexVenusCL is the core communication component of the EaaS system—a GPU P2P communication library built on IBGDA technology, supporting parallel execution of communication and computation, with low latency and high flexibility.
NEX-based agents have demonstrated strong usability and practical value in various real productivity scenarios:
- Agentic Coding: Full-stack mini-program development, web writing
- Academic Research: Poster generation, academic literature analysis
- Deep Research: Comprehensive research reports with multi-source data fusion and multi-modal visualization
- Presentation Production: Automated slide design, material collection, and page production
- And more...
- Project Homepage: https://nex.sii.edu.cn
- GitHub: https://github.com/nex-agi
- Hugging Face: https://huggingface.co/nex-agi
- ModelScope: https://modelscope.cn/organization/nex-agi
- Official Announcement: Shanghai Innovation Institute
Technical report coming soon. Stay tuned!
We look forward to working with developers and users worldwide to build the NEX ecosystem together!
This README was generated using nex-agi/DeepSeek-V3.1-Nex-N1