Skip to content

Ollama (Local)

AgentV supports local model evaluation via Ollama, enabling private, high-speed benchmarking for open-source models like Llama 4.

Use the ollama protocol to connect to a local Ollama instance.

Terminal window
agentv run --path scenarios/loan_scenario.json --protocol ollama --agent ollama://llama4

By default, the harness attempts to connect to http://localhost:11434. You can override this in your .env:

OLLAMA_HOST=http://your-server:11434
OLLAMA_MODEL=llama4

AgentV is optimized for the 2026 local model landscape:

  • llama4 (8B, 70B, 400B)
  • mistral-v4
  • phi-5

A common research use case is comparing a local model against a cloud baseline (e.g., Llama 4 vs. GPT-5.4).

Use the Model Wars protocol:

Terminal window
agentv run --path scenarios/loan_scenario.json --compare agents_inventory.yaml

Local execution speed depends on your hardware (GPU/RAM). If you experience timeouts during evaluation, increase the DEFAULT_ADAPTER_TIMEOUT in your config.