Kimi-K2-Instruct-0905

Kimi K2 Instruct-0905 updates the original Instruct checkpoint. Moonshot extended its context from 128K to 256K and focused the revision on tool-based coding and front-end development.

A changed checkpoint with a longer context

The model card identifies improvements in coding with agents, general coding and the visual quality of generated front-end code. It remains a text generation model: better interface code does not mean that this checkpoint accepts image input.Kimi-K2-Instruct-0905 model card

The coding results include a tool environment

Its coding evaluation documents a modified SWE-agent setup, repeated runs and pruning unreachable Git objects from repositories. These details matter when comparing the published result with another agent framework or with a single attempt.Kimi-K2-Instruct-0905 model card

Keep serving variants separate from model revisions

Kimi K2 attribution is required for commercial products exceeding 100 million monthly users or US$20 million in monthly revenue. The same modified MIT terms apply to the original Instruct and this revision.Kimi K2 Base and Instruct license

Practical guidance. For an API comparison, record the exact endpoint and its model mapping. A faster endpoint can change latency without establishing a new set of trained weights.

Specifications & access

Kimi-K2-Instruct-0905 specifications
DetailValue and conditions
Launch date
AccessDownloadable
Model identifiersmoonshotai/Kimi-K2-Instruct-0905kimi-k2-0905-preview
Context window262,144 tokens

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

Kimi-K2-Instruct-0905 checkpoint configuration
Parameters1 trillion total; 32 billion active per token.Kimi-K2-Instruct-0905 model card
Weight licenseModified MIT.Kimi K2 Base and Instruct license
Former APIkimi-k2-0905-preview is listed as deprecated.Kimi model catalog

Related models

  • Long-reasoning K2 model trained to alternate internal reasoning with repeated tool interactions.

    Launch date Downloadable
  • Original instruction-tuned K2 checkpoint for tool use and coding without extended thinking mode.

    Launch date Downloadable
  • Qwen2.5-72B adaptation trained to locate software faults and produce repository-level fixes.

    Launch date Downloadable