Kimi K3: 2.8 Trillion Parameters and What We Can Still Learn from the Pelican Benchmark

Moonshot AI announced Kimi K3, their most capable model with 2.8 trillion parameters, available now via their website and API. Open-weight release is scheduled for July 27, 2026. Moonshot claims it beats Claude Opus 4.8 and GPT-5.5 on self-reported benchmarks, but loses to Claude Fable 5 and GPT-5.6 Sol. Pricing is $3 per million input tokens and $15 per million output tokens, making it the most expensive Chinese AI model to date. Simon Willison tested K3 with his pelican SVG benchmark via OpenRouter, paying $0.25 for the output. The model used 95 input tokens and 16,658 output tokens (13,241 reasoning tokens). The generated SVG was of poor visual quality but the alt text was good. Willison notes the pelican benchmark is no longer predictive of overall model quality but remains useful for 'hello world' testing, cost estimation, and basic capabilities check.