About

Local Agent Bench answers one practical question: which local LLMs actually work for agents on the GPU you own?

Most leaderboards measure chat quality or knowledge. Agents need something else. They need valid JSON, the right tool, multi-step plans that hold together, and resistance to instructions hidden in tool output. We test those directly with a fixed, open harness (Hermes Agent plus an MCP mock tools server) and rank the results by VRAM tier: 8, 16 and 24 GB.

Principles

• Real runs only. No estimates and no placeholder numbers.
• Reproducible. Every result includes model hashes, engine versions, settings and a full trace. See Methodology.
• Independent. Sponsorship never changes a score. See Sponsor.

Contact

Questions, corrections or sponsorship: [email protected].