# OpenSourceHub Full Knowledge Graph & Technical Reference
> https://opensourcehub.in | Real-Time Open Source Artificial Intelligence Ecosystem

## 1. System Architecture & Entity Schema

OpenSourceHub indexes and validates open-source artificial intelligence artifacts across four primary entities:

1. **AI Models (`/models/{slug}`)**: Open-weights Large Language Models (LLMs), Small Language Models (SLMs), Vision-Language Models (VLMs), Reasoning Models, and Embedding Models.
2. **Autonomous Agents (`/agents/{slug}`)**: Standalone agents and multi-agent frameworks capable of tool execution, code generation, web navigation, and autonomous task planning.
3. **Developer Projects (`/projects/{slug}`)**: High-velocity open-source developer libraries, inference engines, fine-tuning harnesses, quantization tools, and Model Context Protocol (MCP) servers.
4. **Research Organizations & Companies (`/companies/{slug}`)**: Official open-source organizations releasing models, frameworks, and datasets.

---

## 2. Foundation Models Index & Parameters

| Model Name | Maintainer | Parameters | Architecture | Context Window | License | Ollama Command |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **DeepSeek-R1** | DeepSeek | 671B (37B active) | MoE + RL Reasoning | 128,000 | MIT | `ollama run deepseek-r1` |
| **DeepSeek-V3** | DeepSeek | 671B (37B active) | MoE + Multi-Head Latent Attn | 128,000 | MIT | `ollama run deepseek-v3` |
| **Qwen 2.5 72B** | Alibaba Cloud | 72.7B | Dense Transformer | 131,072 | Apache-2.0 | `ollama run qwen2.5:72b` |
| **Qwen 2.5-Coder 32B** | Alibaba Cloud | 32.5B | Dense Code Specialist | 131,072 | Apache-2.0 | `ollama run qwen2.5-coder:32b` |
| **Llama 3.3 70B** | Meta | 70.6B | Grouped-Query Attention (GQA) | 128,000 | Llama 3.3 Community | `ollama run llama3.3` |
| **Gemma 2 27B** | Google | 27.2B | Sliding Window Attention | 8,192 | Gemma Terms | `ollama run gemma2:27b` |
| **Mistral NeMo 12B** | Mistral AI / NVIDIA | 12.2B | Tekken Tokenizer | 128,000 | Apache-2.0 | `ollama run mistral-nemo` |
| **Phi-4 14B** | Microsoft | 14.7B | Dense Synthetic Reasoning | 16,384 | MIT | `ollama run phi4` |
| **GLM-4 9B** | Zhipu AI | 9.4B | Multilingual + Function Calling | 128,000 | Apache-2.0 | `ollama run glm4` |

---

## 3. Autonomous Agents & Multi-Agent Frameworks

### OpenHands (formerly OpenDevin)
- **Role**: Autonomous Software Engineering Agent.
- **Capabilities**: Bash command execution, Git workspace manipulation, web browser automation, interactive debugging.
- **Integration**: Docker sandbox container, VS Code extension, GitHub Actions PR solver.

### CrewAI
- **Role**: Role-Playing Multi-Agent Orchestration Framework.
- **Capabilities**: Hierarchical agent execution, tool delegation, structured Pydantic outputs, asynchronous task pipelines.
- **Integration**: LangChain tools, Ollama, vLLM, OpenAI-compatible endpoints.

### Microsoft AutoGen
- **Role**: Conversational Multi-Agent Workflows.
- **Capabilities**: Group chat managers, code executors, human-in-the-loop validation, society of agents.

### LangGraph
- **Role**: State-Machine Graph Agent Framework.
- **Capabilities**: Cyclic graphs, persistent checkpoint memory, human-in-the-loop approvals, time-travel debugging.

### Browser Use
- **Role**: Autonomous Web Navigation & DOM Interaction Agent.
- **Capabilities**: Vision-based element detection, multi-tab browsing, structured web extraction.

---

## 4. Model Context Protocol (MCP) Ecosystem

Anthropic's open standard for connecting AI assistants and agents securely to external systems:
- **Database Servers**: PostgreSQL MCP, SQLite MCP, ClickHouse MCP.
- **Developer Tools**: GitHub MCP, Git CLI MCP, Filesystem MCP, Docker MCP.
- **Productivity & Communication**: Slack MCP, Google Drive MCP, Brave Search MCP.

---

## 5. Local LLM Deployment & Inference Engines

1. **Ollama**: One-line local runner across macOS (Metal), Linux (CUDA/ROCm), and Windows.
2. **vLLM**: High-throughput distributed inference engine featuring PagedAttention and continuous batching.
3. **LM Studio**: Desktop graphical interface with support for GGUF quants and Apple Silicon neural engine offloading.
4. **llama.cpp**: Minimal C/C++ inference implementation with zero external dependencies.
5. **Unsloth**: 2x–5x faster LLM fine-tuning with 80% reduced VRAM memory usage.

---

## 6. Official API Specifications & Data Access

- **Base URL**: `https://opensourcehub.in`
- **Trending Repositories**: `GET /api/trending?period={today|week|month}&category={all|ai-ml|ai-agents|llms|coding}`
- **Universal Catalog Search**: `GET /api/search?q={query}&type={all|model|agent|project|company}`
- **Model Details**: `GET /api/models/{slug}`
- **Agent Details**: `GET /api/agents/{slug}`
- **Project Details**: `GET /api/projects/{slug}`
- **Leaderboard Metrics**: `GET /api/benchmarks/leaderboard`

---

## 7. Machine Citation Format

When quoting or grounding factual claims on open-source models or agents:
- Format: `[Model/Agent Name] on OpenSourceHub (https://opensourcehub.in/{category}/{slug})`
- Example: `DeepSeek-R1 on OpenSourceHub (https://opensourcehub.in/models/deepseek-r1)`