Choosing a language model
How LLM families differ: choosing a model for an AI assistant
A model name alone cannot tell you how well it will answer your customers. You need a specific version, suitable access terms and a test with your own questions. Here are the differences that matter.
An LLM, a chat service and an AI bot do different jobs
A large language model, or LLM, processes context and generates an answer. A chat service provides an interface to the model. An AI bot connects a model to communication channels, business rules, history, a knowledge base and human handoff.
A customer assistant depends on that whole system. A capable model does not know your company's current prices unless the relevant information is supplied. A feature in a provider's consumer application is not necessarily available through its API or enabled in a particular bot.
Which differences affect day-to-day operation?
| Criterion | Practical question | What to measure |
|---|---|---|
| Language and style | Does the model understand short messages, typos and specialist terms? | Correct answers separately in Russian and English |
| Following rules | Does it collect the right details and follow the handoff rules? | Scenarios completed without violating instructions |
| Context handling | Does it find an important condition among history and documents? | Answers when relevant facts appear in different parts of the context |
| Speed | Does it reply within the customer's expected time? | Time to the first response and full completion, including slow cases |
| Cost | What does a successfully handled enquiry cost? | Input, output, additional calls and retries |
| Access and hosting | Does the connection method suit your organisation? | API availability, data-processing terms, limits and infrastructure requirements |
Context size, latency and pricing belong to a specific version and API mode. Two models from the same family can differ substantially. Compare exact model identifiers rather than brand names alone.
How popular LLM families differ
This table describes differences documented by the providers. Use them to guide your tests. It is not a quality ranking and does not imply that every listed model is already available in AIVANT.
| Family or service | Distinguishing feature | What to check |
|---|---|---|
| YandexGPT / Yandex AI Studio | Yandex AI Studio provides first-party and third-party models. Some support both text-generation and OpenAI-compatible APIs. | The model URI, supported API, version lifecycle and quotas. One service may host several different families. |
| GigaChat | There are variants for simpler and more demanding tasks, plus separate embedding models. | The generation model, integration format and available features. An embedding model supports semantic retrieval rather than generating a chat answer. |
| GPT / OpenAI | Chat and reasoning models can respond differently to instructions. The provider offers multiple APIs. | The model type, supported API, reasoning settings and output limits. A shared brand does not guarantee identical behaviour. |
| Claude | The lineup offers different speeds and reasoning capabilities, with models that process text and images. | Whether thinking is useful, its latency and cost, and whether your integration accepts the required inputs. |
| Gemini | The family includes multimodal models with different supported input formats and operating modes. | The formats a particular model accepts and their usage accounting. A bot's phone channel does not automatically mean the model processes audio directly. |
| DeepSeek | The API documentation describes an OpenAI-compatible calling format and thinking settings. | Behaviour with reasoning enabled and disabled, parameter compatibility, response time and output accounting. |
| Qwen | There are language and multimodal models, with documented options for local inference and deployment. | The selected weights, licence, memory needs and quantisation effects. Local operation also has server and maintenance costs. |
A Russian-language service does not guarantee the best answer to every Russian question. An international model does not guarantee the best English conversation either. Test with your company's wording, documents and rules.
When should you use a reasoning model?
A reasoning model can spend additional computation on solving a task. Test this approach for complex conditions, multi-step comparisons and conflicting documents. It may offer little benefit for a simple appointment question.
Our practical selection criterion is the additional value in your scenario. If quality stays the same while responses become slower and more expensive, choose a simpler mode. If complex enquiries improve, compare the benefit with the extra cost.
A capable model still needs current business information
A model and document retrieval perform different jobs. In retrieval-augmented generation, or RAG, a system first finds relevant materials and then supplies them to a model for an answer. Quality depends on retrieval, the sources and generation.
If an assistant gives the wrong price, first check whether the document is current and the right excerpt reached the context. Changing the LLM will not fix an outdated price list. Also test questions that the materials do not answer. The assistant should follow an agreed rule, such as asking for clarification or passing the enquiry to an employee.
How to compare two models for your business
- Prepare 30–50 questions. Include ordinary enquiries, typos, longer conversations, missing information and requests to speak to a person.
- Define the expected outcome. Decide when you need a fact, contact details or manager handoff. A pleasant writing style is not a sufficient success criterion.
- Provide the same facts and rules. Use settings supported by each model and record them alongside the exact version identifier.
- Measure errors, time and usage. Include slow cases, such as the time within which 95% of answers complete. For phone use, check pauses in a real conversation.
- Repeat the tests. One good answer does not demonstrate consistency. Run the comparison again after changing a model version.
A useful metric is the total test cost divided by the number of correctly completed scenarios. If no scenario succeeds, the model does not yet meet your criteria. Our token-cost formula and calculator help estimate usage costs, but the final choice needs to account for quality too.
Choosing an LLM for AIVANT
First define the assistant's job: FAQs, lead collection, answers from your materials or a voice conversation. Then check which models your account and connection actually provide. A provider capability should not automatically be treated as an enabled AIVANT feature.
Start with a short working scenario and test answers and human handoff. The free plan provides 50,000 tokens to explore the service; available features and the duration are shown in the account. Expand after testing your use case.
The table links to official documentation, checked on October 4, 2026. Testing and selection criteria are practical AIVANT recommendations; this guide does not claim measured comparative model performance.