Chat AI (LLM models)

IntroductionCopy Anchor Link

Chat AI provides a browser interface to AI models hosted by CERIT-SC within e-INFRA CZ. The interface, called Open WebUI, lets you work with text, code, and documents without installing software locally.

You can use it to revise scientific writing, explore programming questions, extract information, and discuss research material. Some models also support images or tools for calculations.

This guide explains how to get access, choose a model, and use the main features. For scripts and automated workflows, see the OpenAI-compatible API guide.

Accessing Open WebUICopy Anchor Link

How to Get Access

Choose the registration route that applies to you.

If you cannot register or sign in, contact k8s@cerit-sc.cz.

Data privacyCopy Anchor Link

Before uploading research material, consider its sensitivity and your project’s data-handling requirements.

The standard hosted models run within e-INFRA CZ. However, external search tools and separately arranged third-party services may send information outside this infrastructure.

The chat interface stores conversations and uploaded files. The AI data privacy policy describes storage, logging, retention, and deletion limitations.

Direct API access bypasses the chat interface’s conversation storage. This does not, by itself, establish that a workflow meets your data-protection requirements.

Logging inCopy Anchor Link

1. Open Chat AI.
2. Select Login.
3. Choose e-INFRA CZ and complete authentication.
4. Once signed in, open a chat.

Using AI ModelsCopy Anchor Link

Start your first conversationCopy Anchor Link

1. Select a model from the menu at the top of the chat. Scroll through the menu to see additional models.
2. Describe your task and provide the context needed to answer it.
3. Specify the desired output, such as a paragraph, table, or code example.
4. Send your message.
5. Review the response and ask follow-up questions where needed.

Choose a model for your taskCopy Anchor Link

Start with a model that meets your needs on a small, representative sample. A larger or slower model does not necessarily produce better results for every task.

Image understanding and image generation are separate capabilities. To create or edit an image, use the image tool described in Examples.

Currently available modelsCopy Anchor Link

Check the live status page for current availability. The tables below describe the documented model inventory; deployments and settings may change.

For programmatic access, use a maintained model alias or an exact model identifier returned by the API. Model identifiers are case-sensitive.

Generative modelsCopy Anchor Link

In the table below:

• Context is the documented context-window size, expressed in tokens.
• Max output is the documented output-token limit.
• Quant. describes the numerical representation used for model weights.
• — means that no value is listed here.

These limits describe capacity, not guaranteed response length, accuracy, or processing speed. Effective limits can also depend on service configuration.

• Kimi K3 — Moonshot AI’s open-weight 2.8T-parameter Mixture-of-Experts (MoE) flagship model. Native multimodal with vision support, a 1M-token context window, and strong performance in long-horizon coding, reasoning, knowledge work, and agentic workflows.
• GLM 5.3 — 756B-parameter language model optimized for advanced reasoning, tool use, and multi-step agentic tasks.
• GLM 5.3 Flash — 320B-parameter (18B active) Mixture-of-Experts multimodal model from ZAI for reasoning, coding, and agentic tasks. The NVIDIA NVFP4 build reduces weight memory by about 3.3 times compared with BF16 and keeps benchmark accuracy close to the BF16 baseline.
• GPT-OSS-120B — OpenAI’s 120B-parameter open-weight language model offering strong general-purpose performance across reasoning, coding, and tool use.
• DeepSeek V4.1 Flash-1017 — efficient reasoning model with a small activated parameter count relative to leading proprietary models. Excels at mathematics, logical reasoning, and code generation.
• Qwen3.8 Flash Next — 180B (MoE) multimodal model with improved coding performance. fp8 quantization, 256k context, and a maximum output of 48k tokens.
• Qwen3.8 27B — 27B (dense) multimodal model with significantly improved coding performance. Supports images and videos, with configurable thinking for balancing speed and reasoning.
• Mistral Medium 3.5 — 128B multimodal model for text, coding, tool use, and agentic workflows.
• Gemma 4 — Google’s 31B multimodal model; useful for OCR/vision and general-purpose tasks.
• Whisper Large v3 — OpenAI’s 1.55B automatic speech recognition (ASR) and speech-translation model; supports 99+ languages with robust transcription even in noisy conditions. Available via API only. See its model card for background information. It is not a text-chat model.

Model aliasesCopy Anchor Link

An alias is a maintained name that points to a deployed model. Its target can change when the service is updated.

Aliases reduce the need to change model names in scripts, but they do not guarantee identical behaviour, outputs, or uninterrupted access.

The documented deepseek and deepseek-thinking targets use the same underlying model weights with different reasoning defaults.

Use an exact model identifier when you need to record which version was used. An exact identifier does not reserve that version or guarantee its future availability.

Embedding modelsCopy Anchor Link

Embedding models represent text as numerical vectors for semantic search and document retrieval. Rerankers perform a different task: they reorder candidate results according to their relevance to a query.

These services are accessed programmatically. See the API guide for access and supported requests.

Check the deployed configuration before designing a document-processing pipeline. Input limits, supported languages, and preprocessing requirements should be evaluated for your dataset.

Programming languages proficiencyCopy Anchor Link

The following results are reported from evaluations using the Aider Polyglot Benchmark.

Use them as supporting information when choosing a model. Performance on benchmark exercises does not establish correctness for your software or research workflow.

How models are added, upgraded, and removedCopy Anchor Link

CERIT-SC evaluates new models using benchmarks, community feedback, and their suitability for the available computing resources.

Because GPU capacity is shared and limited, model versions may be replaced or retired. Maintenance or urgent infrastructure requirements can also temporarily affect availability.

Changes are announced through the documentation and chat-interface notices. Check the live status page when planning a substantial workload.

Preserve your research resultsCopy Anchor Link

A model may change during a thesis, publication, or other research project. For work that needs to be traceable:

• Record the exact model identifier and available metadata.
• Save prompts, relevant inputs, generation settings, and dates.
• Keep copies of outputs in accordance with your data-management plan.
• Re-evaluate your workflow after a model change.

Even with the same model, repeated requests may produce different outputs.

ExamplesCopy Anchor Link

The examples below describe the documented interface. Available controls depend on the selected model and enabled tools.

Revise scientific writingCopy Anchor Link

Provide the passage and specify what may change.

Revise this abstract for clarity and concision. Preserve the scientific meaning, numerical results, citations, and uncertainty. 
Do not introduce new claims. Return the revised abstract followed by any ambiguities that need my attention.

Compare the revision with the original, particularly where it describes results, limitations, or causal relationships.

Explain or debug codeCopy Anchor Link

Include the relevant code, the expected behaviour, and any error message. Remove credentials and other confidential information.

Explain why this Python function fails for an empty input. Suggest the smallest correction and provide examples covering
an empty input, one value, and multiple values.

Review and test generated code before using it in an analysis.

Image-to-text (OCR scanner)Copy Anchor Link

Use a model with image support to extract text from a scanned page or photograph.

1. Select a multimodal model.
2. Select + → Upload Files and upload the image.
3. Describe the extraction task.
4. Send the request and compare the output with the source.

Example:

Transcribe the text in this image. Preserve headings, line breaks, numerical values, and units. 
Mark unreadable text as [unclear] instead of guessing.

Check mathematical symbols, decimal separators, subscripts, and table alignment carefully.

For a scanned PDF, check whether direct upload works with your selected model and tool configuration. Individual page images can be used when necessary.

Image generationCopy Anchor Link

Use the image tool to create illustrations for teaching, communication, or early design work.

1. Select a chat model.
2. Open the Integration control and enable Image.
3. Describe the image you want.
4. Send the request.
5. Refine the result with follow-up instructions.

Example:

Create a simple illustration of a researcher using a notebook, a computing cluster, and a document collection. 
Use a white background, blue accents, and no text.

Generated illustrations should not be presented as experimental observations. Use plotting or diagramming software when values, geometry, or scientific relationships must be exact.

Picture editorCopy Anchor Link

1. Select a model with image support.
2. Select + → Upload Files and upload an image.
3. Open Integration and enable Image.
4. Specify the changes and what should remain unchanged.
5. Send the request and inspect the result.

Example:

Change the background of this conceptual illustration to white.
Preserve the objects, their positions, and their colours.

Keep the original file. Inspect details that the tool may have changed unintentionally.

Python sandbox for math/code tasksCopy Anchor Link

The Python sandbox supports specific calculations and data transformations with text output. It is not a general-purpose programming environment.

1. Select a model that supports tools.
2. Open Integration → Tools.
3. Enable Python Sandbox.
4. Describe the calculation and provide the required data.
5. Inspect the calculation and its output.

Example:

Calculate the mean and sample standard deviation of these measurements: 2.1, 2.4, 2.2, 2.6, and 2.3.
Show the Python code and state which standard-deviation convention you used.

The documented sandbox returns text and does not render plots. For plotting or a reproducible analysis environment, use JupyterHub or your own Python or R setup.

Memorize (personalization)Copy Anchor Link

Memory can retain preferences for later conversations, where supported. For example, you might store that you prefer Python examples or British English.

1. Open your user menu.
2. Go to Settings → Personalization → Memory.
3. Enable memory if it is available.
4. Use Manage to review and edit stored preferences.

Memory support depends on the configured model and interface. Check that a preference has been saved rather than relying only on the model’s acknowledgement.

Do not store passwords, API keys, or other credentials in memory. For reproducible work, include important instructions explicitly in the prompt rather than relying on personalisation.

Knowledge function (alpha)Copy Anchor Link

The experimental Knowledge function lets you query a collection of uploaded documents using retrieval-augmented generation (RAG).

RAG retrieves relevant passages and supplies them as context to a model. Uploading documents does not mean that the model has been trained on them.

The documented setup uses a shared embedding-model configuration. Changing that model can require documents to be processed again.

Retrieval quality depends on the language, document structure, text extraction, and how documents are split into passages. Test representative material before relying on the results.

Check answers against the original passages. Retrieved text and citations do not guarantee that the model’s interpretation is correct.

Before uploading a document collection, review the data privacy policy, including file-retention and deletion limitations.

Creating an API keyCopy Anchor Link

An API key lets your scripts and applications authenticate to the AI service.

1. Open Settings in Chat AI.
2. Select Account.
3. Open API keys.
4. Generate a key or display an existing one. Use the API key rather than the separately displayed JWT token.
5. Store the key securely and configure your application to use it.

The documented API base URL is:

https://llm.ai.e-infra.cz/v1/

Do not share API keys or include them in notebooks, source repositories, screenshots, or chat messages. Use an environment variable or a suitable secret-management mechanism.

For requests, supported features, and client configuration, follow the API guide.

Monitor API usageCopy Anchor Link

The usage dashboard shows token consumption through the API after you sign in with your e-INFRA CZ account.

It does not include token consumption through the chat interface.

A practical guide to using AI chatCopy Anchor Link

A useful prompt usually contains four elements:

1. Task: what you want the model to do.
2. Context: the material and background needed for the task.
3. Constraints: what to preserve, exclude, or treat as uncertain.
4. Output: the required format and level of detail.

For example:

Using only the supplied methods section, extract the sample size,
instruments, experimental conditions, and statistical tests.
Return a table with one row per item and a supporting quotation.
Write “not stated” when the section does not provide the information.

Review the response before refining the prompt. If something is missing or incorrect, identify it directly and explain the correction you need.

Clear instructions can improve usefulness, but they do not guarantee accuracy. Verify references, calculations, extracted values, and scientific claims against appropriate sources.

Shared resources and benchmarkingCopy Anchor Link

The platform is shared across the research community. Response times vary with model choice, request length, and current demand.

For automated work, limit concurrent requests and evaluate your workflow on a small sample before processing a large dataset.

Measurements on a shared service include queueing and other operational effects. Record workload conditions when interpreting performance results.

Help and troubleshootingCopy Anchor Link

For assistance, contact k8s@cerit-sc.cz. Include the approximate time, model name, and a description of the problem. Remove sensitive content and credentials from examples and screenshots.

For community discussion and announcements, see the Matrix communication platform.