Kimi K2.6 by Moonshot AI at lowcloud

Kimi K2.6 is an AI language model from Moonshot AI, a company based in China, released on 20 April 2026. It handles text and images. The context window holds 256,000 tokens; tokens are the pieces a model splits text into, roughly half a word to a whole word each. 256,000 tokens are a few hundred pages of text, give or take, that the model can take in during one pass. The weights, meaning the trained file the model consists of, are public on Hugging Face.

At lowcloud you pick Kimi K2.6 yourself: per message in chat, fixed for one agent, or for a single step of a workflow. The request does not go to Moonshot AI. It goes through Melious AI GmbH in Saarbrücken to servers in the EU. The model also has a fixed second job in the app: if someone attaches an image to a request for a model that cannot read images, lowcloud runs that one request on Kimi K2.6.

Vendor: Moonshot AI, China
Context window: 256k Token
Reads images: yes
Licence: Modified MIT
Processing: EU servers via Melious

Where Kimi K2.6 comes from

The maker is Moonshot AI, a company based in China. The release date is 20 April 2026. Technically it is a Mixture-of-Experts model, MoE for short: the model has around one trillion parameters, but only part of them compute any single request, about 32 billion according to Moonshot. Images are handled by a separate vision part called MoonViT. The weights are public on Hugging Face at moonshotai/Kimi-K2.6. The licence is a Modified MIT License. It is not the plain MIT licence, even though many overview pages simply write "MIT" or "open source". The MIT text is unchanged, and Moonshot added one condition: anyone using the model in a product with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must display "Kimi K2.6" prominently in the product interface. Below those thresholds the usual MIT rights apply. Kimi K2.6 is no longer Moonshot's newest model. Kimi K2.7 Code followed in June 2026 and Kimi K3 in July 2026. That does not make K2.6 retired: Moonshot discontinued the earlier Kimi K2.5 on 31 August 2026, while K2.6 is still in the model list.

What "on EU servers" actually means here

The request never reaches the model maker. The chain is: you or your client, then lowcloud UG in Bielefeld, then Melious AI GmbH in Saarbrücken, then data centres in the EU. Melious does not run its own data centres. It distributes requests across European operators. lowcloud buys capacity there and passes the access on inside the app. Your contract is with lowcloud. What lowcloud does not know, and therefore does not claim: which operator and which data centre serves an individual Kimi K2.6 request. Melious does not publish which operator runs which model. If your client asks, this is the full answer you can give: processing in the EU, contract with lowcloud, service provider Melious AI in Saarbrücken, and the list of individual data centres sits with Melious. The model itself is a file. Moonshot AI being based in China says nothing about where that file is executed.

The fallback model when an image hits a text-only model

Kimi K2.6 has a second role at lowcloud. Not every model in the EU catalogue can read images. If someone attaches an image to a request aimed at a text-only model, the app does not fail. It runs that one request on Kimi K2.6 instead. The text-only models in the EU catalogue are GLM 5.3, GLM 5.2, GLM 5.1, GLM 5, Qwen 3.8 Max, Qwen 3.5 397B, DeepSeek V4 Pro and DeepSeek V4 Flash. The models that read images directly are Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B and Mistral Small 4. What that means for your agents: even with a text model configured, a single request carrying an image can run on Kimi K2.6. The cost of that run then follows Kimi K2.6 pricing and, as always, is shown in euros on the run.

What Kimi K2.6 is good for in the builder

Workflows that involve images. A photo of an invoice, a screenshot from a ticket, a scanned delivery note: the model reads them directly, so no separate text recognition step is needed. Steps that need a fixed output format. A JSON schema is a template that dictates which fields an answer must contain, for example invoice number, date and amount. lowcloud tested against the API and found that Kimi K2.6 accepts such fixed schemas. Moonshot publicly documents only the simpler JSON mode without a fixed schema. That fixed schemas are accepted is lowcloud's own observation, not a vendor statement. The combination of both is currently unique in the EU catalogue. Kimi K2.6 is the only one of the 14 models that reads images and accepts a fixed JSON schema. For a step that has to pull clean fields out of a photo, that is exactly the pairing you need. Longer runs with tool calls. Moonshot describes K2.6 as a model for multi-step agent tasks and published its own benchmark figures in the announcement post of 20 April 2026. Those are vendor figures; lowcloud has not verified them. Choosing the model is part of the free account. Without an account you can have a workflow built at agent.lowcloud.de, and lowcloud picks the model for it.

What Kimi K2.6 is not good for

Very long inputs. 256,000 tokens is the limit. If you want to process a whole contract archive or months of conversation history in one pass, pick a model from the EU catalogue with a 1 million token context: Kimi K3, GLM 5.3, GLM 5.2 or DeepSeek V4 Pro. The newest coding work. Moonshot released two newer models after K2.6: Kimi K2.7 Code in June 2026, explicitly for programming tasks, and Kimi K3 in July 2026. Both are in the EU catalogue as well. Clients who rule out models from Chinese makers as a matter of policy. That requirement does come up, even when processing happens in the EU. In lowcloud's EU catalogue, Mistral Small 4 from Mistral AI in France is the only model not made by a Chinese company. It reads images too, but does not accept fixed JSON schemas.

What Kimi K2.6 costs

There is no figure on this page. A margin is added on top of the purchase price, so any number published here would be wrong. What can be said: within lowcloud's EU catalogue, Kimi K2.6 sits in the lower middle on price. The same catalogue holds cheaper and more expensive models. If price is the deciding factor, K2.6 is not the first pick. Every run shows its cost in euros in the app. For a workflow with several steps on different models, that is where you see what the run actually cost. It is also the number you can bill your client against.

Try Kimi K2.6

Picking a model needs a free account. You then choose Kimi K2.6 in the chat, per agent, or per step.

Book a call

Common questions

What is Kimi K2.6?

An AI language model from Moonshot AI in China, released on 20 April 2026. It handles text and images, and the context window holds 256,000 tokens. The weights are public and the licence is an MIT licence with one added condition. At lowcloud you can select it for an agent or a single step, with processing on servers in the EU.

Does my data go to China if I use Kimi K2.6?

No. The request goes to lowcloud in Bielefeld, from there to Melious AI GmbH in Saarbrücken, and is processed in data centres in the EU. Moonshot AI never receives the request. What comes from China is the model, meaning the trained file, not the place of processing. What lowcloud cannot tell you: which European operator serves an individual request. Melious does not publish that mapping.

Is Kimi K2.6 open source?

The weights are publicly downloadable. The licence is the Modified MIT License: the MIT text plus one condition. Above 100 million monthly active users or 20 million US dollars in monthly revenue, "Kimi K2.6" must be displayed prominently in the product interface. Writing just "MIT licence" leaves that out. For freelancer and agency projects, those thresholds are out of reach.

Can Kimi K2.6 read images?

Yes. lowcloud tested image understanding against the API itself rather than copying the vendor's claim. Kimi K2.6 is also the model the app switches to when an image is sent to a text-only model. Moonshot additionally documents video input; whether that works through the EU access has not been tested by lowcloud.

What does Kimi K2.6 cost at lowcloud?

No figure is given here, because a margin is added on top of the purchase price and a published number would therefore be wrong. In the EU catalogue, Kimi K2.6 sits in the lower middle on price, with cheaper and more expensive models alongside it. The actual cost in euros is shown on every run in the app.

Should I use Kimi K3 instead of Kimi K2.6?

Kimi K3 is Moonshot's newer and larger model with a 1 million token context, and it is available in the EU catalogue too. For very long inputs, K3 is the choice. Kimi K2.6 still makes sense when a step has to read images and return a fixed JSON schema at the same time, which currently only K2.6 can do in the EU catalogue. Moonshot discontinued the earlier K2.5 on 31 August 2026, not K2.6.

Kimi K2.6 is a product name of Moonshot AI. There is no business relationship between lowcloud and Moonshot AI; we run the openly available model through a European provider. Last updated: 7 September 2026.