Kimi K3

Overview. Kimi K3 is Moonshot AI's open-weights flagship, a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, built on a KDA and Gated MLA attention hybrid with a MoonViT-V2 vision tower. Reasoning is always on — it cannot be turned off — and the model supports a 1,048,576-token context window, one of the largest available today. Developers reach for it when a task demands frontier-level reasoning depth across long documents, images, and video, backed by full open weights for self-hosting.

Key strengths. Moonshot's own evaluation table puts Kimi K3 at 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1, and 84.8 on OSWorld-Verified, placing it among frontier-class models on graduate-level science reasoning and long-horizon agentic tasks. Independent measurement from Artificial Analysis at maximum reasoning effort shows a Coding score of 76.2 and an Agentic score of 50, corroborating strength on real coding and tool-use workflows. It also performs well on browsing and document benchmarks (BrowseComp 91.2, OmniDocBench 91.1), reflecting its always-on reasoning trace together with native tool calling and structured output.

When to use it. Use Kimi K3 for reasoning-heavy work — multi-step coding, agentic tool-use, research and browsing tasks, and math or science problems — where the always-on thinking process and million-token context help rather than hinder. It's a strong fit when you need to reason over very large documents, codebases, or long conversation histories in a single call, or when a task mixes text with images. Because reasoning is mandatory and defaults to maximum effort, expect batch-style latency rather than fast conversational turnaround, so it suits asynchronous pipelines better than latency-sensitive chat.

Limitations. Reasoning cannot be disabled and defaults to the highest effort level, so every call carries visible chain-of-thought overhead that adds latency compared with non-reasoning models. Temperature is fixed at 1.0 and cannot be adjusted, and Moonshot has not published a training-data cutoff date. The model also has no audio input or output support.

Sovereignty note. Kimi K3 is created by Moonshot AI and ships as open weights, so teams can self-host it on Sovereign infrastructure for full control over data residency and processing. Available findings don't confirm CLOUD Act exposure or a training-on-input policy for any specific hosted endpoint, so that should be checked per endpoint before relying on it for regulated workloads.