Kimi K2.5

Moonshot's January 2026 1T-parameter native multimodal agent model with Agent Swarm: 32B active, 256K context, modified MIT.

In short. Kimi K2.5 is a very large AI model from Moonshot AI that reads text, images and video and can split a task among many helper agents. The license is the MIT license with one extra sentence that only affects very large products. It needs a server-class GPU node, and Moonshot has since released K2.6, K2.7 Code and K3.

Summary

Moonshot AI released Kimi K2.5 on 27 January 2026. Moonshot built it by continuing the pre-training of Kimi-K2-Base on about 15 trillion mixed visual and text tokens, so it handles images natively rather than through an add-on. It is a mixture of experts with about 1 trillion parameters and 32 billion active per token, 61 layers (one of them dense), 384 experts with 8 chosen per token plus one shared expert, multi-head latent attention, a 256K context and a 400-million-parameter MoonViT vision encoder. The weights are native INT4 and take about 595 GB. K2.5 has an instant mode and a thinking mode, and Moonshot's Agent Swarm, a research preview at launch, lets it split a task among parallel sub-agents.

In Moonshot's own table K2.5 scores 76.8 on SWE-Bench Verified (GPT-5.2 at xhigh effort 80.0, Claude Opus 4.5 80.9, Gemini 3 Pro 76.2), 73.0 on SWE-Bench Multilingual, 50.7 on SWE-Bench Pro, 50.8 on Terminal-Bench 2.0, 85.0 on LiveCodeBench v6, 87.6 on GPQA-Diamond and 50.2 on Humanity's Last Exam with tools. On BrowseComp it scores 60.6, or 74.9 with context management and 78.4 with Agent Swarm. The technical report is on arXiv (2602.02276). All of these numbers are Moonshot's own.

Moonshot recommends vLLM, SGLang or KTransformers with transformers 4.57.1 or later. Its recommended sampling is temperature 1.0 for thinking mode and 0.6 for instant mode, and video input works only through Moonshot's own API. K2.6 followed in April 2026, then K2.7 Code in June and K3, so we list K2.5 as a legacy model. In the 30 days to 11 November 2026, Hugging Face counted about 391,000 downloads. The license is Modified MIT: only a commercial product with over 100 million monthly active users or over US$20 million of monthly revenue must display "Kimi K2.5" in its interface.

Good for

  • Self-hosted multimodal agents that read screenshots, documents and images
  • Studying early Agent Swarm behavior and parallel sub-agent setups
  • Companies that want an almost-MIT license for a trillion-parameter model
  • Reproducing earlier results or comparing against K2.6, K2.7 Code and K3

Watch out for

  • K2.6 is newer and scores higher in Moonshot's own table, and K2.7 Code and K3 came after it
  • It needs a server-class GPU node, since even the INT4 weights are about 595 GB
  • Video input works only through Moonshot's own API for now
  • All benchmark numbers come from Moonshot, and the comparison models are from late 2025

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the K2.5 (INT4) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
K2.5 (INT4) INT4 1.03T 32B active 2.36 TB 1.18 TB 655 GB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

The standard MIT license with one added sentence. Anyone can use, modify and sell the model, and only very large products (over 100 million monthly active users or over US$20 million of monthly revenue) must show the model's name in their interface.

  • Attribution or model name must be shown

More about Modified MIT

Frequently asked questions

What is Kimi K2.5?
Kimi K2.5 is a very large AI model from Moonshot AI that reads text, images and video and can split a task among many helper agents. The license is the MIT license with one extra sentence that only affects very large products. It needs a server-class GPU node, and Moonshot has since released K2.6, K2.7 Code and K3.
Can I use Kimi K2.5 commercially?
Yes, with conditions. Kimi K2.5 is released under the Modified MIT, which allows commercial use but adds extra rules: Attribution or model name must be shown. Read the license text before you ship a product.
Is Kimi K2.5 open source?
Kimi K2.5 is open-weight, not open source in the strict sense. You can download the model files, but the Modified MIT adds conditions that the Open Source Initiative definition does not allow.
Can I run Kimi K2.5 on my own computer?
Not on a desktop. The smallest variant, K2.5 (INT4), needs about 655 GB at 4-bit precision, which means a GPU server with several data-center cards.
When was Kimi K2.5 released?
Kimi K2.5 was released by Moonshot AI on January 27, 2026.
How long a context window does Kimi K2.5 have?
Kimi K2.5 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Kimi K2.6 Moonshot AI 1.03T 32B active 256K Modified MIT (Conditional) 655 GB GPU server
Kimi K2.7 Code Moonshot AI 1.03T 32B active 256K Modified MIT (Conditional) 655 GB GPU server
Kimi K3 Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
MiniMax-M2.5 MiniMax 229B 10B active 192K MiniMax Model (Conditional) 158 GB Big-memory desktop