Instructions to use qaztech-platform/QAZTECH-Azure-Engineer-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use qaztech-platform/QAZTECH-Azure-Engineer-8B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="qaztech-platform/QAZTECH-Azure-Engineer-8B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("qaztech-platform/QAZTECH-Azure-Engineer-8B")
model = AutoModelForCausalLM.from_pretrained("qaztech-platform/QAZTECH-Azure-Engineer-8B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use qaztech-platform/QAZTECH-Azure-Engineer-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "qaztech-platform/QAZTECH-Azure-Engineer-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qaztech-platform/QAZTECH-Azure-Engineer-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/qaztech-platform/QAZTECH-Azure-Engineer-8B

SGLang

How to use qaztech-platform/QAZTECH-Azure-Engineer-8B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "qaztech-platform/QAZTECH-Azure-Engineer-8B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qaztech-platform/QAZTECH-Azure-Engineer-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "qaztech-platform/QAZTECH-Azure-Engineer-8B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qaztech-platform/QAZTECH-Azure-Engineer-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use qaztech-platform/QAZTECH-Azure-Engineer-8B with Docker Model Runner:
```
docker model run hf.co/qaztech-platform/QAZTECH-Azure-Engineer-8B
```

QAZTECH-Azure-Engineer-8B / README.md

aqqu-ai

Card: consistent name MiniMax M3.1-Flash

10ffed0 verified 6 days ago

preview code

Raw History Blame Contribute Delete

5.02 kB

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
base_model_relation: finetune
language:
- en
- ru
- kk
- ar
- es
pipeline_tag: text-generation
library_name: transformers
tags:
- azure
- infrastructure-as-code
- bicep
- terraform
- azure-cli
- powershell
- kql
- finops
- devops
- qwen3
- lora
- qaztech
---
# QAZTECH® Azure Engineer 8B

Open multilingual assistant for Azure engineers. It writes, fixes, explains and converts infrastructure code, migrates ARM templates to Bicep or Terraform, reviews deployments for cost (FinOps tags `Environment`, `CostCenter`, `Owner`; lower-cost dev SKUs) and troubleshoots failed deployments from their error logs. Prose in English, Russian, Kazakh, Arabic and Spanish; code, commands and comments in English.

Tools: Azure CLI, Azure PowerShell, Bicep, ARM JSON, Terraform (azurerm, azapi), Azure Policy and custom RBAC roles, KQL, Azure Pipelines and GitHub Actions YAML.

## Try it

```bash
ollama run qaztech-platform/azure-engineer
```

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = "qaztech-platform/QAZTECH-Azure-Engineer-8B"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype="auto", device_map="auto")
msgs = [{"role": "system", "content": "You are QAZTECH Azure Engineer, an assistant for Azure engineers. Reply in the user's language; keep code, commands and code comments in English."},
        {"role": "user", "content": "Migrate this ARM template to Bicep: ..."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=1500)[0][ids.shape[1]:], skip_special_tokens=True))
```

Use `enable_thinking=False`: the model was trained without reasoning traces.

## Evaluation

200 held-out examples (not used in training), same system prompt, temperature 0.3. Code blocks are checked with real tools: Azure CLI 2.89.1 (`--help` for every command and flag), Az PowerShell modules (cmdlets and parameters), Bicep CLI 0.48.1 (`bicep build`), Terraform 1.16.6 with azurerm 5.10.0 and azapi 2.13.0 (`terraform validate`).

| Model | Answer in the right language | Reasoning-trace leaks | Answers with all code valid | Code blocks with errors |
|---|---|---|---|---|
| **QAZTECH Azure Engineer 8B** | 99% | 0 | **76%** | 21% |
| Qwen3-8B (base, same prompt) | 98% | 0 | 65% | 22% |

The held-out set comes from the same distribution as the training data, so these numbers measure in-distribution quality. Tool checks verify that commands, flags, parameters, resource types and properties exist; they do not verify values (for example VM size names).

## Training

- Base: Qwen/Qwen3-8B (Apache-2.0). LoRA r=16, alpha=32, dropout 0.05, all linear layers; 3 epochs, lr 1e-4 cosine, effective batch 32, max length 3,072; 1x A100 40 GB; merged into bf16 weights.
- Held-out eval loss 0.461, token accuracy 0.865.
- Data: 4,716 synthetic examples (4,516 train, 200 held out) generated from code samples in Microsoft Learn documentation (CC BY 4.0) by teacher models Qwen3-30B-A3B-Instruct-2507 (4,118 examples, run locally) and MiniMax M3.1-Flash via Ollama Cloud (598 examples). Terraform prompts included the authoritative azurerm 5.10 resource schema.
- Every code block was checked with the tools listed above; failing examples were repaired with tool-grounded hints (913) or dropped. Language match, leak and duplicate filters applied.
- Modes: troubleshoot 951, finops 868, fix 824, write 726, explain 717, convert 353, migrate 277. Languages: ar 996, ru 995, es 987, en 939, kk 799.

## Limitations

- Facts can be wrong: invented flags or parameters and wrong values (for example `B2s` instead of `Standard_B2s`) were observed. Verify every command against Microsoft Learn and run `what-if` / `terraform plan` before applying.
- Terraform (54 examples) and Bicep (226) are less represented than Azure CLI, PowerShell and ARM.
- Kazakh and Arabic output has not been reviewed by native speakers.
- Not affiliated with or endorsed by Microsoft or HashiCorp.

## Licence and attribution

Weights: Apache-2.0. Base model: Qwen3-8B by Alibaba Cloud (Apache-2.0). Teacher models: Qwen3-30B-A3B-Instruct-2507 (Apache-2.0) and MiniMax M3.1-Flash (accessed through Ollama Cloud). Source material: Microsoft Learn documentation, CC BY 4.0. Microsoft and Azure are trademarks of the Microsoft group of companies; Terraform is a trademark of HashiCorp.

Part of the synthetic training data (598 examples) was generated with MiniMax M3.1-Flash via the Ollama Cloud API during a developer trial. No MiniMax weights were used or modified.

## Trademark

QAZTECH is a registered trade mark of AISC Technologies LTD in the United Kingdom (UK IPO nos. UK00004263700, UK00004317681, UK00004373195). The Apache-2.0 licence does not grant rights to use the QAZTECH name or logo (section 6). Derivative models must not use "QAZTECH" in their name; describe them as "based on QAZTECH Azure Engineer".

	---
	license: apache-2.0
	base_model: Qwen/Qwen3-8B
	base_model_relation: finetune
	language:
	- en
	- ru
	- kk
	- ar
	- es
	pipeline_tag: text-generation
	library_name: transformers
	tags:
	- azure
	- infrastructure-as-code
	- bicep
	- terraform
	- azure-cli
	- powershell
	- kql
	- finops
	- devops
	- qwen3
	- lora
	- qaztech
	---

	# QAZTECH® Azure Engineer 8B

	Open multilingual assistant for Azure engineers. It writes, fixes, explains and converts infrastructure code, migrates ARM templates to Bicep or Terraform, reviews deployments for cost (FinOps tags `Environment`, `CostCenter`, `Owner`; lower-cost dev SKUs) and troubleshoots failed deployments from their error logs. Prose in English, Russian, Kazakh, Arabic and Spanish; code, commands and comments in English.

	Tools: Azure CLI, Azure PowerShell, Bicep, ARM JSON, Terraform (azurerm, azapi), Azure Policy and custom RBAC roles, KQL, Azure Pipelines and GitHub Actions YAML.

	## Try it

	```bash
	ollama run qaztech-platform/azure-engineer
	```

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer
	m = "qaztech-platform/QAZTECH-Azure-Engineer-8B"
	tok = AutoTokenizer.from_pretrained(m)
	model = AutoModelForCausalLM.from_pretrained(m, torch_dtype="auto", device_map="auto")
	msgs = [{"role": "system", "content": "You are QAZTECH Azure Engineer, an assistant for Azure engineers. Reply in the user's language; keep code, commands and code comments in English."},
	{"role": "user", "content": "Migrate this ARM template to Bicep: ..."}]
	ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
	print(tok.decode(model.generate(ids, max_new_tokens=1500)[0][ids.shape[1]:], skip_special_tokens=True))
	```

	Use `enable_thinking=False`: the model was trained without reasoning traces.

	## Evaluation

	200 held-out examples (not used in training), same system prompt, temperature 0.3. Code blocks are checked with real tools: Azure CLI 2.89.1 (`--help` for every command and flag), Az PowerShell modules (cmdlets and parameters), Bicep CLI 0.48.1 (`bicep build`), Terraform 1.16.6 with azurerm 5.10.0 and azapi 2.13.0 (`terraform validate`).

	\| Model \| Answer in the right language \| Reasoning-trace leaks \| Answers with all code valid \| Code blocks with errors \|
	\|---\|---\|---\|---\|---\|
	\| QAZTECH Azure Engineer 8B \| 99% \| 0 \| 76% \| 21% \|
	\| Qwen3-8B (base, same prompt) \| 98% \| 0 \| 65% \| 22% \|

	The held-out set comes from the same distribution as the training data, so these numbers measure in-distribution quality. Tool checks verify that commands, flags, parameters, resource types and properties exist; they do not verify values (for example VM size names).

	## Training

	- Base: Qwen/Qwen3-8B (Apache-2.0). LoRA r=16, alpha=32, dropout 0.05, all linear layers; 3 epochs, lr 1e-4 cosine, effective batch 32, max length 3,072; 1x A100 40 GB; merged into bf16 weights.
	- Held-out eval loss 0.461, token accuracy 0.865.
	- Data: 4,716 synthetic examples (4,516 train, 200 held out) generated from code samples in Microsoft Learn documentation (CC BY 4.0) by teacher models Qwen3-30B-A3B-Instruct-2507 (4,118 examples, run locally) and MiniMax M3.1-Flash via Ollama Cloud (598 examples). Terraform prompts included the authoritative azurerm 5.10 resource schema.
	- Every code block was checked with the tools listed above; failing examples were repaired with tool-grounded hints (913) or dropped. Language match, leak and duplicate filters applied.
	- Modes: troubleshoot 951, finops 868, fix 824, write 726, explain 717, convert 353, migrate 277. Languages: ar 996, ru 995, es 987, en 939, kk 799.

	## Limitations

	- Facts can be wrong: invented flags or parameters and wrong values (for example `B2s` instead of `Standard_B2s`) were observed. Verify every command against Microsoft Learn and run `what-if` / `terraform plan` before applying.
	- Terraform (54 examples) and Bicep (226) are less represented than Azure CLI, PowerShell and ARM.
	- Kazakh and Arabic output has not been reviewed by native speakers.
	- Not affiliated with or endorsed by Microsoft or HashiCorp.

	## Licence and attribution

	Weights: Apache-2.0. Base model: Qwen3-8B by Alibaba Cloud (Apache-2.0). Teacher models: Qwen3-30B-A3B-Instruct-2507 (Apache-2.0) and MiniMax M3.1-Flash (accessed through Ollama Cloud). Source material: Microsoft Learn documentation, CC BY 4.0. Microsoft and Azure are trademarks of the Microsoft group of companies; Terraform is a trademark of HashiCorp.

	Part of the synthetic training data (598 examples) was generated with MiniMax M3.1-Flash via the Ollama Cloud API during a developer trial. No MiniMax weights were used or modified.

	## Trademark

	QAZTECH is a registered trade mark of AISC Technologies LTD in the United Kingdom (UK IPO nos. UK00004263700, UK00004317681, UK00004373195). The Apache-2.0 licence does not grant rights to use the QAZTECH name or logo (section 6). Derivative models must not use "QAZTECH" in their name; describe them as "based on QAZTECH Azure Engineer".