Kimi Code now has two distinct lanes: K3 and K2.7 Code
The coding product now separates frontier reasoning from a focused coding model and a high-speed variant.
Use K3 for the hardest long-horizon engineering work. Use K2.7 Code for steady day-to-day coding, and its high-speed variant when latency matters more than quota efficiency.
Three model IDs, two model families
Kimi Code exposes K3 as k3, K2.7 Code as kimi-for-coding, and the faster K2.7 variant as kimi-for-coding-highspeed. Model selection is available from the /model command in the CLI.
The high-speed option is documented as roughly five to six times faster while using three times the quota. That makes it a deliberate speed trade, not a free upgrade.
A practical selection rule
Start with K2.7 Code for contained repository work and familiar implementation tasks. Move to K3 when the job requires deep cross-repository reasoning, visual context, or unusually long execution. Choose high-speed only when iteration time is the bottleneck.
Watch the context cache
Kimi Code warns that switching models invalidates the existing context cache. Starting a fresh session before a model change can reduce unnecessary consumption and makes the evaluation cleaner.