Blog Yenisei
Platform news and articles on working with language models: API, RAG, embeddings and reasoning models.
A Yenisei model at a nominal price, available on every plan and in all of our products.
Anthropic released Fable 5.1 on September 1: a million tokens of context, image input, tool calling. What the model can do and which open models in Yenisei suit the same tasks.
The Yenisei catalog includes OpenAI's open models GPT-OSS 120B and 20B: what they are good at, what to use instead of the closed GPT models, and how to move your code from the OpenAI API by changing one line.
September additions to the Yenisei catalog: a large Qwen, models with context over a million tokens, inexpensive models from 1.68 ₽ per 1M tokens, and a coding assistant.
Ox Alpha is a reasoning model with no named author, built for code and agent workflows. It stayed in the catalog only briefly: what it was and what to switch to.
BGE-M3 and Qwen3 Embedding: how they differ, how vector dimensionality matters, and why multilingual support is essential for Russian.
Dozens of models in the catalog, and you have to pick one. A checklist on context, speed, price, and task type, so you stop guessing and overpaying.
Xiaomi has released an open lineup: the Pro version has reasoning, tool calling and a 1M context. We have added both models and explain who they suit.
With a search-grounded model, answer quality depends on how you phrase the question and check the sources. Techniques using Sonar and Yenisei search as examples.
Where money leaks on LLMs and how to get it back. Counting tokens, matching the model to the task, caching, and trimming context. 1 credit = 1 ₽.
Sonar answers not from memory but from fresh search results with links. How this approach differs from a regular chat model and how to get the same in Yenisei.
Reasoning models solve what ordinary ones break on, but they cost more and take longer. How to pick the effort level and not pay for needless thinking.
An enterprise assistant needs more than a "correct" answer: it needs a reference to the source. We build RAG on Command A so that every statement is verifiable.
Cohere's flagship with an enterprise slant. We explain why Command A is first and foremost a RAG model, and what to take instead of the compact Command R7B.
A model with reasoning and function calling is a convenient foundation for an agent. We build a minimal tool-calling loop on the Yenisei base_url and pick a model for it.
Call transcription, voice bots, meeting analysis. Voxtral and MiMo on audio: how to pass a file, handle long recordings, and keep the meaning intact.
Llama 3.3 70B was one of the most popular open models for a long time. What it was valued for and which Yenisei models offer the same today.
Meta's open lineup: Scout and Maverick understand images, and Scout's window exceeds a million tokens. Which one suits whom.
Reasoning, vision, tools and a 500K context: where Grok 4.6 is strong and which open Yenisei models cover the same tasks.
Mistral has a dedicated model for audio. Voxtral Small transcribes and analyzes speech. We show how to build it into call and voice note processing.
A large window is not free. We show how to build requests to million-token-context models on long documents so that you pay for context deliberately.
Vision models read scans, tables, and screenshots. How to pass an image, extract structure, and avoid overpaying for image tokens.
Mistral models in Yenisei: Small 3.2 with images and a 256K context, the compact Nemo, and Voxtral for audio. Which one for which task, and what to use instead of the departed Large and Medium.
From the cheap text-only Micro to Premier with image input and a million-token context. We walk through the Nova ladder and the open Yenisei models on the same rungs.
K2.7 Code is a dedicated Kimi build for development: where a specialized code model beats a general-purpose one and how to fit it into your workflow.
response_format and JSON Schema make the model answer in a given structure. How to get structured data without regular expressions.
Kimi K2 Thinking unfolds a chain of thought, a regular model answers directly. When reasoning pays off and when it just burns tokens.
Three well-known families of closed models and the open models that cover the same tasks in Yenisei. A practical scenario-by-scenario breakdown, and how not to overpay.
Google's open lineup that reads images. Why an open model when closed flagships are right there, and how to choose between Gemma 4 26B and 31B.
Moonshot's new generation has arrived in the catalog. Kimi K3 combines 1M context, reasoning, image understanding and tool calling. Here is what makes it stand out.
LLMs translate better than classic engines if you set the task properly. Glossaries, tone, preserving markup, and choosing a model for Russian.
Two Google models with a million-token window that understand text, images, audio, and video. Their niche, and the open models in Yenisei with the same set of inputs.
Not every task deserves a flagship. The lightweight Z.AI model handles a large stream of requests cheaply and quickly: classification, routing, short answers.
Next to flagships and reasoning models, a fast model for text and images is easy to write off, and that would be a mistake. Its niche, using GPT-4o as the example, and the open models in Yenisei that fill it.
Two Z.AI models for agents and reasoning. One reasons more thoroughly, the other answers faster. The choice depends on who is waiting for the answer: a person or a program.
Model-written text you are not embarrassed to publish: controlling tone, structure, and length without rewriting by hand.
Reasoning models think longer for the sake of accuracy. How they differ from regular chat models, using o3 and o4-mini as examples, and which ones are available in Yenisei.
The senior Z.AI model of the previous release: 1M context, reasoning and tools. Where it is strong, where it is overkill, and how it compares to GLM 5.3.
The top model of the GPT lineup, with a context of about a million tokens. When a large window really changes things, and when you are paying for a window you don't use.
A mid-size general-purpose reasoning model from Alibaba. When 27B is the right size and how Qwen3.8 differs from the previous Qwen3.5.
How to build a first-line bot that answers in a fraction of a second without breaking the bank. Flash/Mini models, streaming, escalation to a human.
Low latency, 1M context, and image support in a single model, and the products it suits.
Three OpenAI models on one architecture: which GPT-5 to use for which workload, and which open models in Yenisei cover the same tiers.
The top Qwen with a 262K-token window, function calling, and confident Russian: a middle ground between regular chat and a million-token context, and what replaced it.
Anthropic's model focused on code generation and long agent chains. How it differs, and which open models in Yenisei cover the same scenarios.
The specialized coding Qwen: why a narrow model for code beats a general-purpose one and where it is useful in development.
Qwen Coder, Kimi K2.7 Code, DeepSeek: how they differ on real code. How to plug one into your editor without going broke on autocomplete.
The top Qwen had a million-token window and image support. We look at what that gives you in practice and which Qwen3.8 models can do it today.
Anthropic's smallest model is for tasks where speed and price matter more than depth: its role, and the open models in Yenisei that handle the same stream of requests.
Pro, Flash, Flash Vision, or V3.2: each has its own job. A short decision map with examples.
Anthropic's flagship and its balanced smaller sibling. How they differ in practice, and which open models in Yenisei cover the same tasks.
The model only asks to call your function; you are the one who runs it. The tools schema, the execution loop, and the usual pitfalls with parallel calls.
V3.2 is a solid everyday workhorse: around 160K context, function calling, and confident Russian. A good default choice for most products.
R1 showed not only the answer but how it got there. Why seeing the solution path matters and how to get it from the models in Yenisei.
Vector search on bge-m3, POST /v1/embeddings, and context assembly. A working setup, plus the places where RAG usually breaks.
The smaller V4 responds quickly and costs several times less than the flagship. Here are the scenarios where Flash beats Pro on more than just price.
We added DeepSeek's flagship for reasoning and code to the catalog. Here is what you pay for and when it is worth it.