Overview
Kimi K3 is Moonshot AI's flagship model, announced on July 16, 2026. Available through the API as kimi-k3, it combines a sparse Mixture-of-Experts architecture, native image and video understanding, always-on reasoning, and a 1,048,576-token context window for long-horizon coding, knowledge work, and agentic workflows.
#Key Features
- Sparse MoE Architecture: Uses Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and Stable LatentMoE, activating a subset of experts per request for efficient large-scale inference.
- Long-Horizon Agentic Work: Designed to navigate large repositories, coordinate terminal tools, conduct multi-step research, and produce complete software or knowledge-work artifacts with limited supervision.
- Native Visual Understanding: Accepts text, image, and video input and can use visual feedback in frontend engineering, game development, CAD, document analysis, animation, and video-editing workflows.
- 1M Context and Large Outputs: Provides a 1,048,576-token context window with automatic context caching.
max_completion_tokensdefaults to131,072and can be configured up to1,048,576, subject to the combined context limit. - Production API Controls: Supports streaming, function calling,
tool_choice, dynamic tool loading, strict JSON Schema output, Partial Mode, and official Formula tools through an OpenAI-compatible Chat Completions API.
#Best Use Cases
- Large-Repository Software Engineering: Long-running implementation, debugging, repository navigation, terminal orchestration, GPU kernel optimization, compiler work, and visually guided frontend or game development.
- End-to-End Knowledge Work: Research, financial and scientific analysis, document synthesis, spreadsheet tasks, presentation creation, and interactive reports that combine evidence, code, and visualizations.
- Multimodal Agent Workflows: Tasks that require text, screenshots, images, video, tools, and iterative visual feedback within the same agent loop.
- Long-Context Applications: Codebase assistants, fixed-corpus analysis, document Q&A, and agents that benefit from a stable repeated prefix and automatic cache hits.
#Capabilities and Limitations
| Capability | Description |
|---|---|
| Reasoning | Always uses thinking mode; reasoning_effort currently supports only max. |
| Coding | Suitable for long-running coding, repository navigation, terminal workflows, compiler work, GPU kernel optimization, debugging, and visually guided frontend or game development. |
| Creative Writing | Supports general and long-form text generation. |
| Multimodal | Text, image, and video input with text output. |
| Response Speed | Latency depends on reasoning length, output length, media inputs, and tool use. |
| Context Window | 1,048,576 tokens, with automatic caching for repeated initial context. |
| Max Output | 131,072 tokens by default; configurable up to 1,048,576 tokens, subject to the combined context-window limit. |
| Tool Use | Function calling, tool_choice, dynamic tool loading, structured output, JSON Mode, Partial Mode, and official Formula tools. |
| Multilingual | Supports Chinese and English usage; provider-side language coverage may evolve with the Kimi API. |
Usage example
cURL
curl https://www.namoclaw.ai/api/v1/chat/completions \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "moonshotai/kimi-k3-20260715",
"messages": [{ "role": "user", "content": "Hello!" }]
}'Default parameters
| Name | Type | Default | Description |
|---|---|---|---|
| temperature | number | — | — |
| top_p | number | — | — |
| top_k | number | — | — |
| frequency_penalty | number | — | — |
| presence_penalty | number | — | — |
| repetition_penalty | number | — | — |
No reviews yet — be the first to share your experience.