# Kimi K2.5 — weights, license and what is actually open

A trillion-parameter visual agentic model with a comparatively small active footprint and public weights and code.

Canonical: https://brightaifuture.com/open-models/kimi
Format: model-family
Source publication: Not established
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. No independent reproduction is attached; benchmark statements remain attributed to Moonshot AI.

## Documented release

Kimi-K2.5

Organization: Moonshot AI

Kimi K2.5 release: 2026-01-27 (day precision). Source: https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md

Kimi K2.6 release noted by Moonshot AI: 2026-04-20 (day precision). Source: https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md

K2.5 remains the reference checkpoint here; Moonshot AI’s cited documentation notes a later K2.6 release.

Parameters: 1T total / 32B active, plus a 400M vision encoder

Architecture: Multimodal mixture-of-experts transformer with MoonViT vision encoding.

Modalities: text, image, video

Context: 256K tokens

## License and commercial use

Modified MIT License

Generally permitted, with an additional display requirement for products above the license’s stated revenue or user thresholds.

## What is open

Weights: available. Official model weights are downloadable. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Architecture: available. Configuration and architecture details are public. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Inference code: available. The official repository provides code and deployment examples. Source: https://github.com/MoonshotAI/Kimi-K2.5

Training code: not-established. Complete training code was not confirmed. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Training recipe: partial. The paper describes training and agentic methods without a complete reproducible recipe. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Data information: partial. Training strategy is described without release of the full mixture. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Training data: not-established. No sufficiently specific public artifact was confirmed in this review. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Evaluation: partial. Vendor evaluation protocols and results are published. Source: https://huggingface.co/moonshotai/Kimi-K2.5

Commercial use: restricted. The modified MIT license adds a display requirement at defined scale thresholds. Source: https://huggingface.co/moonshotai/Kimi-K2.5

## Uses and strengths described in sources

Visual agent workflows

Coding

Long context

Sparse activation

## Hardware and quantization

The full checkpoint is server-class; quantized serving still requires careful multi-device planning.

Official/community NVFP4, MXFP4, FP8, GGUF, and MLX variants are listed on the Hub

## Independent evidence

No independent reproduction is attached; benchmark statements remain attributed to Moonshot AI.

## Limitations

Extremely large total checkpoint

Custom license condition at large commercial scale

Training data is not released

## Provenance and history

{}

## Original sources

- [Moonshot AI Kimi K2.5 release note](https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md)
- [Kimi K2.5 official repository](https://github.com/MoonshotAI/Kimi-K2.5)
- [Kimi K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)
- [Kimi K2.5: Visual Agentic Intelligence](https://arxiv.org/abs/2602.02276)

## Continue exploring

- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence thread](https://brightaifuture.com/threads/open)