The Amazon Nova lineup: which model is for whom and what replaces it in Yenisei

The Amazon Nova family is four models covering the range from high-volume cheap requests to heavy tasks with images and long context. Amazon has built a clear ladder where each rung has an obvious purpose.

There are no Nova models in the Yenisei catalog: we work with open-weight models and our own Yenisei lineup. But the ladder itself is a good way to think about model choice, and we have a model of our own for every rung.

Who is who in Nova

• Nova Premier is the senior model: context up to 1M tokens, image input, tool calling.
• Nova Pro and Nova Lite are models with image input and a 300K context; Pro is smarter and more expensive, Lite is faster and cheaper.
• Nova Micro is text-only and the cheapest: classification, field extraction, short answers.

The same ladder in Yenisei

• Cheap text stream (like Micro): Mistral Nemo (mistral-nemo), GPT-OSS 20B (gpt-oss-20b), Granite 4.1 8B (granite-8b).
• Fast, with image input (like Lite): Yenisei 100 (yenisei-100), Mistral Small 3.2 (mistral-small-3.2), Gemma 4 26B (gemma-4-26b).
• Stronger, with image input (like Pro): Qwen3.8 Flash (qwen-flash), Llama 4 Maverick (llama-4-maverick), Gemma 4 31B (gemma-4-31b).
• Flagship with a million-token window (like Premier): Yenisei Max (yenisei-max), Kimi K3 (kimi-k3), Llama 4 Scout (llama-4-scout); the last one has a window over a million tokens.

How to call

All models live at the standard OpenAI-compatible address:

curl https://api.yenisei.ru/v1/chat/completions \
  -H "Authorization: Bearer $YENISEI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-nemo",
    "messages": [
      {"role": "user", "content": "Determine the sentiment of the review: positive, negative or neutral. One word only."}
    ]
  }'

Change model to yenisei-max and the same code goes to the flagship.

How to choose

Start from the bottom. If the task boils down to a short answer and there will be a lot of them, take a lightweight text model: it is cheap and answers quickly. Need an image in the request but at moderate volume: a mid-range model with image input. And only when context really becomes the bottleneck (a million tokens is hundreds of pages) or you need maximum accuracy on a hard task: the flagship.

A typical mistake is to put the flagship in right away "just to be safe". On a stream of simple requests this turns into a bill that is unpleasant to explain later.

Prices for all models are on the plans and pricing page, billed in rubles, 1 credit = 1 ₽.