A million tokens and vision in Qwen: from Qwen3.7 Plus to Qwen3.8

Qwen3.7 Plus was the top model of Alibaba's Qwen lineup, with a 1M-token context. In the Yenisei catalog it has been replaced by the Qwen3.8 generation: Qwen3.8 Flash (qwen-flash), with a million tokens, images and video as input, and reasoning; and Qwen3.8 2.4T (qwen3.8-2.4t), the largest model of the lineup, with a million-token window for text. With a million-token context you can solve tasks that don't fit into a regular window.

What they can do

• 1M-token context. An entire book, the whole repository of a mid-size project, or a year of correspondence fits in the window.
• Images. Qwen3.8 Flash accepts screenshots, diagrams, scanned documents, charts, and video.
• Tools. Full function calling in both.
• Reasoning. Both can reason before answering.
• High-quality Russian. Writes cleanly, without calques.

Why a million tokens

With a large context you don't need workarounds: no slicing the document into chunks, no clever search over them, no stitching answers together. You put in the whole material at once, and the model sees connections that get lost when the text is chopped up.

Example: analyzing a large contract. If it is chopped up, the model won't notice that clause 14 contradicts appendix B when they land in different chunks. With a million-token context, it will.

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-2.4t",
    "messages": [
      {"role": "user", "content": "Here is the entire contract. Find internal contradictions between its sections."}
    ]
  }'

Images in practice

With images you can, for example, parse an interface screenshot and describe a bug, extract a table from a scan, or explain an architecture diagram. All through the same chat/completions endpoint with an image in the message, using the qwen-flash model.

When the million-token window is overkill

A million-token context costs money and time: the model processes the entire window. If your requests fit into 30-40K tokens, the huge window is unnecessary, and a lighter model will answer cheaper and faster. Look at Qwen3.8 Flash (also 1M, but faster and lighter) or at Qwen3.8 27B (qwen3.8-27b) if 262K is enough.

Use the large model when you really do put a great deal into the context and need an answer across all of the material at once.