What is Kimi K3? Architecture, context window, vision, and agentic capabilities

01

Kimi K3 in one clear definition

Kimi K3 is Moonshot AI's open-weight, native multimodal model for long-horizon coding, knowledge work, tool use, and reasoning. The official model card describes a 2.8-trillion-parameter mixture-of-experts model with 104 billion activated parameters per token, native image understanding, and a 1,048,576-token context window.

Searchers sometimes call it “Kimi V3,” but Moonshot's official product and repository use Kimi K3. That distinction matters when looking for documentation, API model names, weights, pricing, and reliable comparisons.

02

How the sparse mixture-of-experts architecture works

A mixture-of-experts model contains many specialist parameter blocks and routes each token through a smaller subset. K3 selects 16 of 896 routed experts, plus shared experts, instead of activating all 2.8 trillion parameters for every token. This does not make the model small: the official card still lists 104 billion active parameters per token.

The practical purpose is to increase model capacity without paying the full dense-compute cost on every step. Real speed and cost still depend on serving hardware, quantization, batching, context length, cache behavior, and the number of reasoning and output tokens produced.

• 93 model layers, including one dense layer
• 896 routed experts with 16 selected per token
• 104B activated parameters per token
• MXFP4 weights and MXFP8 activations

03

Kimi Delta Attention and long context

K3 combines Kimi Delta Attention with gated multi-head latent attention. Moonshot positions this architecture as a more efficient way to scale attention and memory across very long inputs. The one-million-token limit can accommodate large repositories or document collections, but maximum context is not the same as maximum useful evidence.

Long prompts can increase latency, cost, distraction, and the chance that weak evidence receives too much weight. A strong application still retrieves the smallest relevant evidence set, labels sources, separates instructions from documents, and tests whether additional context improves the answer.

04

Native vision and agentic work

The model card lists text and image as supported modalities and identifies MoonViT-V2 as the vision encoder. Moonshot presents K3 as a model for coding sessions, repository navigation, terminal tools, research, dashboards, and other multi-step work.

These are system-level tasks. The model needs safe tool definitions, permissions, state, budgets, error handling, and human approval around consequential actions. A benchmark result or large context window does not replace production controls.

05

Where Kimi K3 fits—and where it does not

K3 is most interesting for teams evaluating open weights, very long context, multimodal inputs, coding, and agentic workflows. It is not automatically the best choice for short classification, high-volume extraction, low-latency chat, or work already handled reliably by a smaller model.

Evaluate it with a representative task set against alternatives. Measure completion quality, citation accuracy, tool-call success, latency, cache hit rate, total output tokens, and cost per accepted result. Treat vendor benchmarks as useful evidence, not a substitute for your own test.