Tip
At Microsoft Build 2026 (June 2, 2026), Microsoft unveiled MAI-Thinking-1 — their first in-house reasoning model, now available in private preview on Azure AI Foundry. If you are evaluating reasoning-tier models for your enterprise applications, this one is worth putting on your shortlist.
What Makes MAI-Thinking-1 Different?
Most reasoning models available today were trained using knowledge distillation from other large models. MAI-Thinking-1 was trained from scratch with zero distillation, using only commercially licensed data — which matters significantly if you are building in regulated industries where training data provenance is auditable.
Key specs:
| Attribute | Value |
|---|---|
| Active parameters | 35 billion |
| Context window | 256K tokens |
| Training data | Zero distillation — commercially licensed only |
| SWE Bench Pro | Benchmarked to match Anthropic Claude Opus 4.6 |
| Blind preference tests | Preferred over Claude Sonnet 4.6 |
How to Access It
MAI-Thinking-1 is currently in private preview on Azure AI Foundry. To request access:
- Go to Azure AI Foundry
- Navigate to Model Catalog
- Search for MAI-Thinking-1
- Select Request Access and complete the preview form
Once approved, you can call it via the standard Azure AI Inference SDK:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
from azure.ai.inference import ChatCompletionsClient
from azure.core.credentials import AzureKeyCredential
client = ChatCompletionsClient(
endpoint="https://<your-endpoint>.models.ai.azure.com",
credential=AzureKeyCredential("<your-key>"),
)
response = client.complete(
model="mai-thinking-1",
messages=[
{
"role": "user",
"content": "Walk through the reasoning for why a circular buffer is a good fit for a fixed-size log store."
}
],
max_tokens=4096,
)
print(response.choices[0].message.content)
When to Use a Reasoning Model
Reasoning models like MAI-Thinking-1 shine for tasks that benefit from deliberate, step-by-step thinking:
- Complex code generation — multi-file refactors, algorithm design
- Document analysis — legal contracts, technical specifications with interdependencies
- Multi-hop Q&A — questions that require combining facts from multiple sources
- Evaluation and planning — designing test plans, architecture reviews
For simple chat or retrieval-augmented generation over well-structured data, a standard model will still be faster and more cost-efficient.
The Broader MAI Model Family
MAI-Thinking-1 is part of a larger family of 7 models Microsoft announced at Build 2026:
| Model | Use case |
|---|---|
| MAI-Thinking-1 | Complex reasoning tasks |
| MAI-Image-2.5 + Flash | Text-to-image and image-to-image generation |
| MAI Transcribe 1.5 | Speech-to-text, 43 languages |
| MAI-Voice-2 + Flash | Text-to-speech, 15+ additional languages |
| MAI-Code-1 | Inference-efficient coding model for GitHub/VS Code |