Kimi-K2.5

Kimi K2.5 continues training from K2 Base on mixed visual and text data. It combines instant and thinking behavior within one checkpoint and adds a MoonViT vision encoder.

Visual input changes the coding workflow

The launch article demonstrates using images and video for interface generation and visual debugging. The training combines visual and text tokens rather than placing a separate image-captioning service in front of an unchanged K2 text model.Kimi K2.5 visual reasoning and agent training

Product modes do not create four model checkpoints

At launch, Kimi offered Instant, Thinking, Agent and Agent Swarm modes. The swarm system used a trained orchestrator with parallel workers. These product modes and their surrounding tools should not be counted as four independently released weight models.Kimi K2.5 visual reasoning and agent training

Use the K2.5 input format

The model card records corrections to its default system prompt and image-start token after release. A local application needs the matching multimodal processor and chat template.Kimi-K2.5 model card

The license requires Kimi K2.5 attribution for products above 100 million monthly users or US$20 million monthly revenue.Kimi K2.5 license

Specifications & access

Kimi-K2.5 specifications
DetailValue and conditions
Launch date
AccessDownloadable
Model identifiersmoonshotai/Kimi-K2.5kimi-k2.5
Context window262,144 tokens

Published checkpoint limit; this is not a promise of continued API access.

Kimi-K2.5 checkpoint configuration
Parameters1 trillion total; 32 billion active. The vision encoder is listed separately as 400 million parameters.Kimi-K2.5 model card
Training ancestryKimi-K2-Base, followed by approximately 15 trillion visual and text tokens.Kimi-K2.5 model card
Weight licenseModified MIT.Kimi K2.5 license
Former APIkimi-k2.5 is listed as deprecated; weights remain published.Kimi model catalog

Related models

  • A text, image and video model with selectable thinking and non-thinking modes.

    Launch date not confirmedDeprecated API; check access detailsAPI · Downloadable
  • Long-reasoning K2 model trained to alternate internal reasoning with repeated tool interactions.

    Launch date Downloadable
  • Trillion-parameter pretrained model that introduced MuonClip training to the Kimi K2 family.

    Launch date July 2025Downloadable