Moonshot: Kimi K2 tool-call accuracy varies up to 16 points across providers

Moonshot's K2 Vendor Verifier repo compares 15 providers serving Kimi K2 Thinking on the same 4,000-request test set: schema_accuracy ranges from 83.05% at Chutes to 100% at MoonshotAI and Fireworks. On K2 0905 the spread reaches 28 points, from 100% down to 71.96% at Together.

• Kimi K2 Thinking: 83.05% (Chutes) to 100% (MoonshotAI, Fireworks) at temperature 1.0
• K2 0905: five providers at 100%, Together at 71.96%, Baseten at 72.49%
• Groq scored 100% schema_accuracy but only 69.52% tool-call trigger F1
• At 87.22% per-call accuracy, a 25-call session validates just 3.3% of the time