A 27B reasoning model from Qwen: Qwen3.8 27B in the catalog

The catalog has a mid-size general-purpose reasoning model, Qwen3.8 27B (qwen3.8-27b). When this article first came out, Qwen3.5 27B was alongside it; it is no longer in the catalog. Qwen3.8 27B occupies a convenient niche between light chat models and heavy flagships, and on top of that it understands images and video.

Why 27B is often the right size

In many products, a 27B model turns out to be a good compromise. It is large enough for reasoning: breaking a task into steps, not skipping a link in the logic, handling a multi-step problem. At the same time it is compact enough to respond quickly and cost a reasonable amount.

You don't always need a flagship. If your tasks are "think it through and justify" rather than "solve an olympiad problem," 27B covers them with room to spare, and you don't overpay for capacity you don't use.

How 3.8 differs from the earlier 3.5

• Qwen3.5 27B was proven and stable.
• Qwen3.8 27B is the newer version: more careful in reasoning and in Russian, better at holding long chains.

If your system was tuned for 3.5, run your typical requests through 3.8 and compare the answers before migrating.

Example

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b",
    "messages": [
      {"role": "user", "content": "Compare two cache architecture options and justify which one to choose for write-heavy workloads."}
    ]
  }'

When to use Qwen3.8 27B and when to use others

• You need mid-depth reasoning at a reasonable price: Qwen3.8 27B.
• You need maximum quality on the hardest tasks: the flagship, DeepSeek V4 Pro (deepseek-pro).
• You need explicit reasoning steps to verify the logic: Kimi K2 Thinking (kimi-k2-thinking).
• You need a huge context: Qwen3.8 Flash (qwen-flash) or Qwen3.8 2.4T (qwen3.8-2.4t).

Qwen3.8 27B's Russian is good; switching to another model is a one-line change in the model field.