Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weights model — the first in the 3T-parameter class — designed for long-horizon coding, knowledge work, and deep reasoning. It introduces Kimi Delta Attention (KDA) for 1M-token context, Attention Residuals, and the Stable LatentMoE architecture activating 16 of 896 experts per token. The guide covers how to use it on Together AI's OpenAI-compatible API: configuring reasoning effort (low/high/max), streaming, vision input, structured output, tool calling with dynamic tool loading, 1M context with automatic prefix caching, and fixed sampling parameters. Pricing is $0.30/M cache-hit input tokens, $3.00/M cache-miss input tokens, and $15.00/M output tokens (thinking tokens billed as output). Benchmark results show K3 leads on SWE Marathon, BrowseComp, and AutomationBench, competing at the GPT-5.6 Sol and Claude Fable 5 tier while being fully open-weights.
