ORBIT Lab is a lightweight, local-first AI models comparision benchmarking focused on open-source language models.
The project aims to simplify the workflow of AI developers by providing a simple and minimal way to compare multiple local models.
New to ORBIT? Start here:
→ Getting Started Guide - Step-by-step setup for beginners
Want technical details?
→ Chat Interface Documentation - Chat architecture, SSE streaming, and system prompts
→ Arena Benchmark Documentation - Benchmarking architecture, metrics extraction, and real-time visualization
| Document | Purpose |
|---|---|
| GETTING_STARTED.md | Complete setup guide with all dependencies and commands |
| ChatInterface.md | Technical architecture of the unified chat interface |
| ArenaBenchmark.md | Details on the multi-model benchmarking pipeline |
| README.md | Project overview and vision (this file) |
- Real-time SSE Streaming: Live chunk-by-chunk markdown rendering.
- Think Mode (Chain of Thought): Expandable thought process blocks with dynamic token counters for reasoning models.
- Stop Generation: Graceful stream abortion using AbortControllers.
- System Prompts & Parameters: Fine-tune context length, temperature, top-p, and set custom personas.
- Performance Metrics: View native token metrics (Tokens/sec, Time-To-First-Token) immediately after generation.
- Session Persistence: Drafts, configurations, and chat histories are seamlessly saved to
localStorage.
- Multi-Model Benchmarking: Queue up to 6 local Ollama models simultaneously.
- Sequential Execution: Models are benchmarked one-by-one to preserve system resources and hardware safety.
- Live Chunk Streaming: Watch each model generate its response in real-time, throttled for UI performance.
- Real-time Throughput: Line Chart tracking tokens/sec across the duration of the entire benchmark generation.
- Throughput Comparison: Vertical Bar Chart directly comparing total speed per model.
- Performance Footprint: Visual Radar Chart normalizing Speed, Latency, Efficiency, and Volume.
- Timeline Breakdown: Stacked Composed Chart detailing Load Time, TTFT, and Generation Time.
- Framework: React 19 + TypeScript + Vite
- Routing: React Router DOM (with Keep-Alive architecture)
- Data Visualization: Recharts
- Markdown Processing: React-Markdown + Remark-GFM + Rehype-Katex
- Icons: Lucide React
- Styling: Vanilla CSS Modules (CSS Variables & Flexbox/Grid)
- Framework: FastAPI (Python)
- AI Integration: Official Ollama Python Client
- Architecture: Modular API router system
- Validation: Pydantic
🚧 Beta Phase
ORBIT has completed its Phase 2 expansion. Both the Chat Workspace and the Model Arena are fully operational with shared streaming utilities, global state persistence, and rich data visualization.
ORBIT/
├── Documents_For_Developers/ # Technical Documentation
├── backend/ # Python FastAPI Backend
│ ├── modules/ # Endpoint routers (chat/, arena/)
│ └── utils/ # Shared logic (llm.py)
├── frontend/ # React UI
│ ├── src/
│ │ ├── components/ # Feature-based folder structure
│ │ │ ├── chat/
│ │ │ ├── arena/
│ │ │ └── common/
│ │ ├── pages/ # React Router pages
│ │ ├── services/ # API fetchers (chatService, arenaService)
│ │ └── types/ # TypeScript Definitions
├── main.py # Uvicorn entry point
└── README.md
Last Updated: June 2026