Real-World Agent Benchmark Compares MiniCPM5-2B and Spark-X2.5-4B Performance

A practical benchmark tested MiniCPM5-2B and Spark-X2.5-4B on a multi-step customer service agent task requiring real database validation and function calling. MiniCPM5-2B completed the entire workflow flawlessly in 8.35 seconds, scoring 100/100, whereas Spark-X2.5-4B scored 84/100 in 12.60 seconds because it stopped to ask the user an unnecessary question instead of completing the pickup schedule. Standard benchmarks often fail to measure 'last-mile agency'—whether a model can execute an autonomous workflow to completion without redundantly passing decision-making back to the user. This benchmark demonstrates that a smaller 2B model can outperform a larger 4B model in both speed and agentic completion on practical tasks. The evaluation required models to verify order details, logic rules, stock at alternative warehouses, set accurate pickup dates, and commit records directly to a backend database. MiniCPM5-2B resolved the task using 527 completion tokens, while Spark-X2.5-4B required 1,460 tokens and failed to execute the final pickup scheduling tool call.