Ubicloud took Kimi's tool-call accuracy from 92% to 100% by flipping one vLLM flag
Fixing Tool-Call Accuracy in vLLM
Serving Kimi K2.6 with vLLM 0.28, Ubicloud saw only 92% of tool calls match the expected schema — two failures in every 25. The culprit wasn't the model but the inference engine's defaults. After enabling vLLM's constrained decoding, which masks invalid tokens during generation using XGrammar structural tags, they hit 100% schema accuracy on the K2VV ToolCall benchmark and shipped it to production.
The quality you observe isn't based on the model alone. The inference engine also plays an important role. Change the inference engine and its settings, and you change the model's output quality.