The Kimi model map
Model IDs, context windows, input modes, reasoning controls, and a practical guide to choosing the right lane.
| Model | Model ID | Context | Input | Reasoning | Status |
|---|---|---|---|---|---|
| Kimi K3 Flagship | kimi-k3 | 1,048,576 | Text + vision | Low / High / Max | Current |
| Kimi K2.7 Code Focused coding | kimi-k2.7-code | 262,144 | Text | Thinking | Current |
| K2.7 Code HighSpeed Fast coding | kimi-k2.7-code-highspeed | 262,144 | Text | Thinking | Current |
| Kimi K2.6 General multimodal | kimi-k2.6 | 262,144 | Text + vision | Thinking / Instant | Supported |
Choose by workload
The largest model is not automatically the most efficient model for every coding session.
Deep, long, or visual work
Use K3 when the task spans a large codebase, combines images with engineering, needs extended reasoning, or must run for a long time with little supervision.
Steady repository work
Use the focused coding model for daily implementation, debugging, and contained agent loops where predictable instruction following matters.
Latency is the bottleneck
Use the high-speed coding model for rapid iteration. Kimi documents it as roughly five to six times faster with three times the quota usage.
The original K2 API model IDs are retired. The current API documentation also gives K2.5 and Moonshot V1 a platform-wide August 31 retirement date following the K3 launch.
What the context number does not tell you
A larger context window changes what can fit in one request, but it does not eliminate the need for retrieval design, acceptance tests, or cost controls. Long-context applications should track cache-hit behavior and task completion together.