Overview
Kimi K3 is an open-weight, natively vision-capable model designed for long-horizon coding, knowledge work, and reasoning. It has 2.8 trillion total parameters and a mixture-of-experts architecture that activates 16 of 896 experts, using Kimi Delta Attention and Attention Residuals. Its context window is 1,048,576 tokens. (kimi.com)
Capabilities and uses
- Understands visual inputs and supports long-context tasks involving large codebases and documents.
- Can assist with software engineering, research, data analysis, interactive visualizations, and agent workflows using tools.
- Is available through Kimi products and the Kimi API; the API model ID is
kimi-k3. (kimi.com)
Limitations
Moonshot reports that output quality can become unstable if an agent framework does not preserve the required reasoning history or switches to Kimi K3 mid-conversation. The model may also make unintended decisions when instructions are ambiguous, so applications should set clear behavioral boundaries and review consequential outputs. Moonshot states that overall performance still trails the strongest proprietary models in its comparison. Self-hosting is demanding: the launch announcement recommends a supernode with at least 64 accelerators. (kimi.com)
Release and availability
The model was released on July 16, 2026. Moonshot announced the release of its weights and technical report on July 27, 2026. The Chinese launch announcement is dated July 17, 2026. (kimi.com)
Official API pricing
Prices below are for kimi-k3, per 1,000,000 tokens. They are separate regional offerings, not currency conversions. Sources were checked on November 16, 2026; confirm current rates and conditions on the relevant platform before use.
- Mainland China, CNY: cache-miss input ¥20; cache-hit input ¥2; output ¥100. An official pricing comparison also lists cache writes at ¥20 for a 5-minute TTL and ¥40 for a 1-hour TTL. That comparison says its rates were checked on October 24, 2026. Sources: Chinese launch announcement and official domestic pricing comparison. (kimi.com)
- International, USD: cache-miss input $3.00; cache-hit input $0.30; output $15.00. Source: English launch announcement. (kimi.com)
Cache-hit prices apply to eligible reused input; cache-write conditions and applicable taxes should be checked for the relevant platform. API usage is billed separately from product subscriptions. (kimi.com)