An AI tutor is useful when it helps a learner ask better questions, practise administrative decisions and verify the answer. QAZTECH AZ-104 Tutor 8B is an open model intended for that kind of work with Microsoft Azure. You can run a quantised version locally, inspect the reproduction pipeline and connect study sessions to official documentation. The aim is a controllable learning tool, not a promise that fluent text is always technically correct.
The QAZTECH AZ-104 repository publishes the pipeline, model links, results and limitations. Its preview release supports English, Russian, Kazakh, Arabic and Spanish. It is based on Qwen3-8B and released under Apache-2.0 for code and weights. The public material explicitly distinguishes the model from Microsoft’s certification programme and does not contain real Microsoft exam questions.
What the model is designed to do
Use the tutor to explain a concept, correct a misconception or create a certification-style practice question with a worked explanation. These are different modes of study. An explanation should help you build a mental model. A misconception exercise should expose a mistaken assumption. A practice question should make you choose between plausible actions and explain why the other options fail. Asking for all three in one large response often makes it harder to check the result.
The repository reports training on a synthetic dataset and provides evaluation information. Those results describe particular tests; they do not establish full AZ-104 readiness or universal multilingual quality. Coverage is currently concentrated on RBAC, virtual networks, Azure Resource Manager and App Service documentation. Storage, monitoring and other exam areas still require official learning resources and independent practice. Native-speaker review of Kazakh and Arabic remains a stated limitation.
Run the Ollama version
After installing Ollama for your operating system, open a terminal and run:
ollama run qaztech-platform/az104-tutorThe published Ollama entry uses a Q4_K_M quantisation. The first run downloads the model, so allow time and disk space for that step. A local installation can then answer without sending each inference request to a hosted model service, provided you actually select the local model and keep your surrounding tools local. Downloading, updating and opening documentation are separate network activities. The complete application determines the privacy boundary, not the word local in its marketing.
Consult Ollama’s local FAQ for memory, context and server configuration. Model-file size alone is not the full memory requirement: runtime buffers and context consume additional memory. Close unnecessary applications and start with short prompts if you are testing on a constrained device. Check that the process is using the intended hardware. Slow generation is a reason to inspect the configuration, rather than silently changing to a remote service.
Choose Transformers or GGUF deliberately
The Hugging Face Transformers release provides the full model. The separate GGUF release provides quantised files for compatible local runtimes. Use the model card and runtime documentation to choose the right file and template. A GGUF file is not automatically interchangeable with a Transformers checkpoint, and a filename does not establish that a runtime supports the architecture.
The repository also links Kaggle variants and ModelScope distribution. Start from the repository’s current links so you do not accidentally follow an older namespace. Record the model identifier, revision and quantisation used in a study session. If an answer changes after an update, that record lets you reproduce the earlier environment and compare results meaningfully.
Use a question format that exposes mistakes
A useful prompt states your level, the topic, the desired response and the verification step. For example: Explain the difference between an Azure subscription and a resource group for a beginner. Give one administrative example, then ask me one question. Identify anything I should verify in Microsoft Learn. This is more useful than requesting an entire certification course in a single response, because you can check one concept before moving on.
For practice, ask the tutor to wait for your answer. Request four options, one correct choice and an explanation after your attempt. Then challenge the explanation: Which permission, scope or service behaviour makes this option correct? Find the relevant documentation yourself. If the answer is uncertain, keep the uncertainty visible. A good study habit is to distinguish what you now understand from what remains a claim requiring verification.
Connect the tutor to a real administrative task
Start with a disposable resource group in a training subscription. Before executing a command, ask what it changes, which account it uses and how you will remove the resource afterwards. Inspect the command rather than copying it immediately. Confirm the active subscription and resource group in Azure. After the operation, check the result independently in the portal or with a read-only command. The tutor’s statement that an operation succeeded is not evidence that Azure completed it.
For a storage exercise, examine the region, redundancy, public-access setting and authentication method before creating anything. Do not paste production secrets, personal documents or privileged credentials into prompts. Explain errors using redacted messages and a minimal reproduction. Cloud resources can incur charges, including in accounts that have a free allowance or promotional credit. A small exercise needs a cleanup step and a way to verify that cleanup.
Study the official scope, not just the model’s strengths
The official AZ-104 study guide is the reference for the exam’s current objectives. Build a checklist from it and mark topics as read, practised and independently checked. The model can help you revisit mistakes, but it should not decide that an entire domain is complete because it answered a few questions fluently. Combine explanations with hands-on work and the official learning material.
Our AZ-104 with an Open AI Tutor — Starter Track introduces setup, Azure concepts, a first disposable lab and a study plan. It is a starter course, not an official Microsoft certification or a guarantee of passing the exam. The practical outcome is a repeatable learning workflow: ask, inspect, execute carefully, verify and record what you learned. For organisations, the same discipline helps turn AI assistance into a controlled deployment conversation.