Kimi K2 and Kimi K2.5 are two generations of Moonshot AI's open model, released six months apart. They are the same size, but Kimi K2.5 learned to see: it understands images and video as well as text. This page compares them in plain words and explains where each stands today.
Kimi K2.5 kept what Kimi K2 could do and added sight and teamwork.
Both have about 1 trillion parameters. Kimi K2.5 was trained further on top of Kimi K2.
It adds images, video, two thinking modes and Agent Swarm, with higher scores.
Kimi K2.5, unless you must match an older Kimi K2 setup.
| Kimi K2 | Kimi K2.5 | |
|---|---|---|
| Released | July 2025 | January 2026 |
| Understands | Text | Text, images and video |
| Reads at once | 128K tokens, later 256K | 256K tokens |
| Thinking | A separate "Thinking" version | Thinking and instant modes in one model |
| Agent Swarm | No | Yes, introduced here |
| Size | 1 trillion parameters, 32 billion active | The same, plus a 400 million vision part |
| Coding test score | 65.8% to 71.3% | 76.8% |
| License | Modified MIT | Modified MIT |
| Moonshot's API | Retired 25 May 2026 | Retired 31 August 2026 |
| Open weights | Still available | Still available |
On a small screen, swipe the table sideways to see every column.
From Moonshot's model cards on Hugging Face and its models page. The coding score is SWE-bench Verified, as published by Moonshot.
Kimi K2.5 was trained on pictures and text together, so it can read screenshots, charts, scanned pages and video.
Because it can see, Kimi K2.5 can turn a screenshot or mockup into working code. Moonshot calls this "coding with vision".
Kimi K2 needed a separate version for step-by-step reasoning. Kimi K2.5 has both built in, and you switch between them.
Kimi K2.5 can split a big job between many helper agents working at once. Earlier research for this site put it at up to 100 helpers.
Moonshot reported higher scores for Kimi K2.5 on coding, reasoning and web research tests. See the next section for the coding figures.
These ideas live on in today's tools: Visual Coding · Design to Code · Agent Swarm
SWE-bench Verified asks a model to fix real bugs in real software projects. The score is the share it fixes.
All four figures are Moonshot's own, from its model cards. We did not run these tests. The versions were tested at different times with different setups, so read this as a trend and not an exact measurement.
Kimi K2.5 also has scores that Kimi K2 cannot have at all, because they test images and video. Moonshot reported 92.3 on OCRBench, a test of reading text in pictures.
About 1 trillion parameters, with 32 billion used for each piece of text. Both need the same class of server hardware.
Both are free to download under the same Modified MIT License.
Both were built for coding and for tasks where the model uses tools by itself.
Answer two questions for a suggestion.
| Where | Kimi K2 | Kimi K2.5 |
|---|---|---|
| Kimi app | No Replaced by newer models | No Replaced by newer models |
| Moonshot's API | Retired 25 May 2026 | Retired 31 August 2026 |
| Your own servers | Yes Weights on Hugging Face | Yes Weights on Hugging Face |
| Other hosts | Varies Check the provider | Varies Check the provider |
On a small screen, swipe the table sideways to see every column.
Still calling kimi-k2.5 or a kimi-k2 name in your code? Those requests no longer work on Moonshot's API. See moving off Kimi K2 and Kimi API setup.
If you do not need these exact models, a current one will serve you better. Moonshot recommends Kimi K3 as the replacement for both.