Ubicloud turns on vLLM constrained decoding to fix tool-call accuracy

Ubicloud enabled constrained decoding in vLLM to improve tool-call schema accuracy for two models, Kimi K2.6 and DeepSeek V4 Flash. The report says accuracy for both rose from 92% to 100%. That schema accuracy is the only result confirmed so far; the report gives no figures for other models or other accuracy metrics.

Written by AI from the coverage · updated 1 hour ago

Not enough continuous observations to draw a trend yet.