Kimi K3
Kimi K3 is the preceding Kimi flagship, available through the API and as model weights. Its API keeps reasoning enabled and offers a million-token context.
A large window still needs a budget
The API defaults max_completion_tokens to 131,072 and permits a value up to 1,048,576. Moonshot requires the complete assistant message to be returned for later turns.Kimi K3 API guide
Practical guidance. Choose an output budget for the task. A large permitted maximum is not a reason to request a book-length response to a short question. Keep space for the input documents and the continuing conversation.
The cache lifetime cannot change on a hit
K3 caching offers five-minute and one-hour tiers. The first write fixes the tier; later hits renew it. A different tier takes effect only after the existing entry expires and the prefix is written again.Kimi context caching
Practical guidance. If people pause to review a draft between turns, measure those gaps before choosing a cache lifetime. A setting that works for a continuous agent loop can behave differently in a human review workflow.
Separate bulk work from interactive work
Kimi batch inference is priced at 60% of the standard rate. Batch cache hits are best effort, and unfinished jobs expire at the selected completion window.Kimi batch billing
Practical guidance. Use batch processing when a delayed result is acceptable. Keep interactive requests on a path that meets the user’s response-time needs.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | kimi-k3moonshotai/Kimi-K3 |
| Context window | 1,048,576 tokens Published API context limit. Input, reasoning and output must fit within the request budget. Kimi inference billing |
| Reasoning controls | Low, high and max; max is the default. Reasoning cannot be disabled.Kimi K3 API guide |
| Sampling controls | Temperature, top-p, sample count and penalties are fixed in the API.Kimi K3 API guide |
| Weight license | Kimi K3 License.Kimi K3 license |
Related models
- Kimi K3.1 Pro
Moonshot AI · Kimi K3.1
A K3-derived multimodal model with sparse attention, 2.8T total parameters and 104B active parameters.
Launch date not confirmedAPI · Downloadable - Kimi K2.6
Moonshot AI · Kimi K2.6
A text, image and video model with selectable thinking and non-thinking modes.
Launch date not confirmedDeprecated API; check access detailsAPI · Downloadable