How to Run Reliable Local LLM Agents on an RTX 3090: A Benchmark (5 Models, Priced in Watts)

Short summary

Benchmarked 5 open-weight LLMs on a single RTX 3090 with two agent orchestrators—opencode and native tool-calling LangGraph. Key finding: tool-adherence (whether models call tools vs chat) was the master variable; GLM-4.5-Air went 0% to 93% accuracy with better orchestration. Qwen3-Coder 30B won overall on speed, correctness, and per-task energy cost.

• •Framework mattered more than model: GLM-4.5-Air jumped 0% → 93% when switched from opencode to LangGraph
• •Tool-adherence is the master variable; native tool-calling fixes models that opencode breaks
• •Qwen3-Coder 30B emerged as all-weather winner: fast, accurate, energy-efficient via MoE

Generated with AI, which can make mistakes.