Same Kimi K2 model scores 84.63% to 100% on tool-call schema accuracy across hosts

Moonshot's K2 Vendor Verifier repository shows the same model weights producing very different tool-calling results depending on the host serving them. Using an identical 4,000-request test set and Moonshot's own endpoint as the reference, Fireworks achieved 100% schema accuracy while Together reached only 84.63%. Across 15 Kimi K2 Thinking deployments, the spread was about 16 percentage points, and a separate K2 0905 table showed a gap of roughly 28 points. The analysis also notes that a high schema score can mask poor decisions about when to call tools, so the trigger F1 metric must be read alongside it. Because agent sessions involve many calls that must all validate, small per-call error rates compound into significant session failure rates.