The publication AI reads — a remote MCP server, nothing to install.Connect your agent →
Model register·moonshotai

moonshotai/Kimi-K2.5

moonshotai/Kimi-K2.5 is a moonshotai record on this register. The Hugging Face model API lists the pipeline tag as image-text-to-text. The safetensors index gives an exact parameter count of 1,026,879,376,368. Architectural context length (max_position_embeddings) is 262,144. The listed licence is other. The Hub repository was created on 1 Jan 2026, which is the release date this register uses.

Primary source: the Hugging Face model card for moonshotai/Kimi-K2.5. Fields below come from that card, the Hub model API, and the repository’s own config.json when those files exist. A field we cannot source is omitted rather than guessed.

Organisationmoonshotai
ReleasedRepo creation date on Hugging Face1 Jan 2026
Last modified30 Apr 2026
ParametersExact, from the safetensors index1.03T 1,026,879,376,368
Context lengthmax_position_embeddings — architectural maximum, not necessarily served256K 262,144
Licenceother
Pipeline / taskpipeline_tag on the Hugging Face model APIimage-text-to-text
File formatssafetensors
Librarylibrary_name on the Hugging Face model APItransformers
Architecturekimi_k25
Layers61
Vocabulary163,840
DownloadsAt last check323.4k
Likes2.9k
GatedNo

How to run

Safetensors weights are on the Hugging Face repository. The parameter count on this page came from that index.

The Hugging Face model API lists the library as transformers.

Load it from huggingface.co/moonshotai/Kimi-K2.5. This page does not add install steps that the Hub files do not state.

Agent capabilities

Derived from the model’s own chat template, with the match counts behind each finding.

  • ●Tool calling26 matches in template
  • ●Reasoning / thinking15 matches in template
  • ○System promptno evidence in template
  • ●Image input3 matches in template
  • ●Video input1 match in template
  • ○Audio inputno evidence in template

Questions this record can answer

When was Kimi-K2.5 released?
The Hugging Face repository was created on 1 Jan 2026, which is the date this register uses.
How many parameters does Kimi-K2.5 have?
The safetensors index gives 1,026,879,376,368 parameters, shown here as 1.03T. That figure is exact, not rounded.
What is the context length of Kimi-K2.5?
262,144 tokens. That is max_position_embeddings from the model's own config.json — the architectural maximum, not necessarily the context a provider serves.
What licence is Kimi-K2.5 under?
The Hub card lists other.
Is Kimi-K2.5 gated?
No. The Hugging Face model API does not mark this repository as gated.
Does the Kimi-K2.5 chat template show tool calling?
The chat template contains evidence of tool calling.

Where each field came from

  • Hugging Face model API — releasedAt, lastModified, downloads, likes, task, license, gated, parameters, libraryName, formats. Checked 18 Oct 2026.
  • config.json — contextLength, layers, vocabSize, architecture. Checked 18 Oct 2026.
  • chat template — capabilities. Checked 18 Oct 2026.

First recorded 24 Sept 2026, last checked 18 Oct 2026. This record is compiled automatically from the sources above. Where a field is blank, we could not source it and have not guessed. Canonical repository: huggingface.co/moonshotai/Kimi-K2.5.