Kimi K3 from Moonshot: a million tokens, vision and tools

We have added Kimi K3 (kimi-k3), the new generation of Moonshot models. The main difference from earlier Kimi models: reasoning, image understanding and tool calling are now in one model, and the context has grown to 1M tokens.

What it can do

• 1M context. Moonshot has always been known for long context, and here it is honest: there is no case where the spec claims 1M but the model gets confused after 200K. It stays coherent across long documents.
• Images. It understands diagrams, interface screenshots, tables in pictures, charts. Working through documentation with illustrations gets noticeably easier.
• Reasoning and tools. It reasons before answering and calls functions confidently.

A request with an image

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "What is wrong with this architecture diagram?"},
        {"type": "image_url", "image_url": {"url": "https://example.com/diagram.png"}}
      ]
    }]
  }'

Russian language

Kimi has traditionally been slightly better at Russian than many Chinese models, a side effect of its focus on long texts. K3 writes coherently and keeps its style over long distances. This shows when analyzing a large Russian document: the model does not "forget" the terminology by the end.

Who it is for

Kimi K3 is a strong choice when long context and images come together in one task: parsing PDFs with diagrams, analyzing design mockups alongside the spec, working with scanned documents. If you do not need images, only reasoning, compare it with Kimi K2 Thinking, which can be more cost-effective. For code, look at the dedicated Kimi K2.7 Code.

The model is in the catalog, accessible with a regular key.