models and data
Which models run and where your data goes
ChatNeuron does not train its own models: the assistant and AI employees use provider models or your own. Below: which, where they run and what reaches them.
- 01Three pathsRussian models (Sber), foreign ones via the AITunnel intermediary, your own model on your server.
- 02Masking before the modelPhones, email, names, passport, tax ID and SNILS are replaced with tokens.
- 03Russian mode for personal data“Russian models only” switch, or your own model with no cloud fallback.
models
Models the platform works with
The account owner picks the model in the employee settings. The default model is currently a foreign one — we say so up front.
| Model | Vendor | Origin | How we call it | Status |
|---|---|---|---|---|
| GigaChat 2 Lite | Sber | Russian | Directly to the Sber API | Selectable |
| GigaChat 2 Pro and Max | Sber | Russian | Directly to the Sber API | Connected; not shown in the employee selector |
| YandexGPT Pro and Lite | Yandex | Russian | Directly to Yandex Cloud | Connected; act as fallback in “Russian models only” mode |
| GPT-5.6 Luna (default), GPT-6 Luna, GPT-4.1 Mini | OpenAI | Foreign | Via the AITunnel intermediary | Selectable |
| DeepSeek Chat | DeepSeek | Foreign | Via the AITunnel intermediary | Selectable |
| Your own model | Yours | On your server | Compatible server: Ollama, vLLM, LM Studio, llama.cpp | Connected in the dashboard |
GigaChat 2 Lite
- Vendor
- Sber
- Origin
- Russian
- How we call it
- Directly to the Sber API
- Status
- Selectable
GigaChat 2 Pro and Max
- Vendor
- Sber
- Origin
- Russian
- How we call it
- Directly to the Sber API
- Status
- Connected; not shown in the employee selector
YandexGPT Pro and Lite
- Vendor
- Yandex
- Origin
- Russian
- How we call it
- Directly to Yandex Cloud
- Status
- Connected; act as fallback in “Russian models only” mode
GPT-5.6 Luna (default), GPT-6 Luna, GPT-4.1 Mini
- Vendor
- OpenAI
- Origin
- Foreign
- How we call it
- Via the AITunnel intermediary
- Status
- Selectable
DeepSeek Chat
- Vendor
- DeepSeek
- Origin
- Foreign
- How we call it
- Via the AITunnel intermediary
- Status
- Selectable
Your own model
- Vendor
- Yours
- Origin
- On your server
- How we call it
- Compatible server: Ollama, vLLM, LM Studio, llama.cpp
- Status
- Connected in the dashboard
The selectable list is limited by plan economics: it holds models that fit the cost of a dialog. Other supported models are used only by internal platform scenarios.
what leaves
What the model receives and what stays with you
We never send a copy of your systems to a model. Only what is needed to answer the specific request leaves.
Goes to the model
- the request text after masking
- knowledge-base fragments and system data needed for the answer and visible to the employee
- employee instructions and the current dialog history
Stays with us and with you
- a copy of your CRM, email and accounting system — we do not make one
- access keys to your systems — stored encrypted and never given to a model
- the action log: who, when, what was called
- listed personal data: phones, email, names, passport, tax ID, SNILS, birth dates, licences, address — replaced with tokens
ChatNeuron servers and the database are in the IT Park data centre in Kazan. The model itself runs at its provider, not on our servers: Russian ones at Russian providers, foreign ones via an intermediary at their developers — outside Russia. Masking recognises data by format and cannot guarantee hiding of arbitrary text.
model choice
How a task reaches the right model
Within the chosen set the platform decides whether a request needs a light or a strong model. A rule decides, not a separate model call.
Light and strong
Light: classification, extraction, short answers, summaries. Strong: multi-step actions, planning, an explicit request for detailed analysis.
By signals in the request
We look at the wording (“compare”, “draw up a plan”), the number of steps and systems, how many tools are needed. The same request gets the same decision.
Free and instant
There is no separate model call for routing: it costs nothing and adds no delay.
Adjustable threshold
A company can move the threshold at which the strong model switches on.
Website chat — the owner's model
In chat for your customers, the model chosen by the employee's owner answers.
Fallback models
If a model does not answer, the next one in the chain is used. In “Russian models only” mode and with your own model the chain never leaves them.
russian mode
“Russian models only” switch
One switch for the whole account. While it is on, requests are processed only by Russian models.
Default model — GigaChat 2
Working requests with tools use GigaChat; YandexGPT acts as fallback when connected.
No silent switching
If the Russian model is unavailable, the assistant refuses. The request does not go to a foreign model.
All surfaces
Applies to customer-facing AI employees, the assistant inside systems and the dashboard helper.
Where to enable
During sign-up in the questionnaire, or later in Settings → AI models. You can turn it off the same way at any time.
own model
Own model and the box: data stays inside your perimeter
If an account has its own model, all its answers go only there, with no cloud fallbacks.
Discuss the closed perimeterYour server only
A request goes neither to cloud models nor to third-party Russian ones. If your model does not answer, you get an error, not a silent move to the cloud.
Own embeddings
Knowledge-base search can be moved to your server (for example, bge-m3).
The box
In the closed perimeter the administrator sets the model and cloud calls are off.
in progress
What we are still clarifying
What is missing or being confirmed is not presented as done.
AITunnel legal entity and processing country
Being clarified and will be added to the data-processing policy. For personal data use the Russian mode or your own model for now.
Roskomnadzor notification on cross-border transfer
Status is shown in the Trust Center. Until confirmed, do not send personal data of Russian citizens through foreign models.
We will show it on your scenario
We will connect a test system, pick a mode and show in the log what reached the model.