Kimi K3
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 to 131,072 and can be configured up to 1,048,576, 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. |
Known Limitations
- Kimi K3 currently supports only
reasoning_effort="max";temperature,top_p,n, presence penalty, and frequency penalty are fixed by the API. - Multi-turn and tool-calling requests must return the complete assistant message, including thinking history. Omitting that history or switching to Kimi K3 midway through a session can make output quality unstable.
- Applications with strict boundaries should provide explicit constraints in the system prompt or
AGENTS.md. - Public image URLs are not accepted directly. Images must use base64 or an
ms://file reference, and videos must be uploaded before use. - Built-in web-search support is not listed in the current B.AI pricing table for this model. Check the platform model catalog for the latest runtime availability.
Standard Pricing
The token prices below are shown in USD per 1 million tokens; web search is billed in USD per use.
| Model | Input (USD / 1M Tokens) | Cache Write (USD / 1M Tokens) | Cache Read (USD / 1M Tokens) | Output (USD / 1M Tokens) | Web Search (USD / use) | Billing Notes |
|---|---|---|---|---|---|---|
| Kimi K3 | $3.00 | $3.00 | $0.30 | $15.00 | - | Cache write is billed at the same rate as input; cache read is billed at 0.1x input |
Credits settlement: B.AI converts charges at 1 USD = 1,000,000 Credits and deducts Credits from the account balance.
Pricing note
Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through account benefits. Specific prices and account benefits are subject to the platform display and final billing records.