Moonshot AI has launched Kimi K3, touting it as their "most capable model to date, with 2.8 trillion parameters." Currently accessible via web and API, an open-weight release is slated for July 27, 2026. Moonshot claims Kimi K3 is the first "open 3T-class model," surpassing Deep Seek's 1.6T v4 Pro. Benchmarks suggest K3 often outperforms Claude Opus 4.8 and GPT-5.5 high, though it trails Claude Fable 5 and GPT-5.6 Sol.

Kimi K3 demonstrates a significant improvement in token usage, reducing output tokens by 21% compared to Kimi K2.6. It also leads on Arena.ai’s Frontend Code arena. However, its pricing—$3/million input tokens and $15/million output tokens—makes it the most expensive model from a Chinese AI lab to date, a substantial increase from previous versions.

Despite its advancements, the article uses a long-standing "pelican benchmark" to illustrate limitations. The prompt "Generate an SVG of a pelican riding a bicycle," while revealing Kimi K3's high reasoning token usage and cost for a simple task, highlights that such benchmarks are less effective for comparing advanced models, especially regarding agentic tool calling and long-context reliability. The benchmark's value now lies in serving as a basic "hello world" for testing new models and confirming fundamental capabilities.