Skip to main content

Local & Offline Models

AgentX features first-class support for zero-cost, privacy-first local inference. When using local models, all code generation, planning, and tool calls remain 100% on your machine.


Supported Runners

  • Ollama (Recommended for macOS, Linux, and Windows)
  • LM Studio (GUI runner with visual model download manager)
  • llama.cpp (High performance CLI server)
  • Jan / LocalAI (Open source local inference servers)

ModelMemory FootprintRecommended Use Case
Qwen 2.5 Coder 7B~5 GB VRAMBest overall coding agent balance
Qwen 2.5 Coder 14B~10 GB VRAMHigh-accuracy architecture & multi-file editing
DeepSeek-R1 Distill 8B~6 GB VRAMDeep reasoning, bug diagnostics, and algorithms
Llama 3.2 3B~3 GB VRAMFast lightweight scripts and rapid iteration

Automated Ollama Setup

AgentX can automatically detect your hardware specs (CPU threads, total RAM, and GPU VRAM) to recommend the optimal quantization level:

# Pull recommended model
ollama run qwen2.5-coder:7b

# Launch AgentX with Ollama
agentx --provider ollama