---
description: Model card for Qwen3.6-27B: a 27B multimodal (vision + text) model with thinking and non-thinking modes, flagship-level agentic coding, tool use, and near-instant responses on Groq.
title: Qwen 3.6 27B - GroqDocs
image: https://console.groq.com/og_cloudv5.jpg
---

Loading model information...

### [Key Technical Specifications](#key-technical-specifications)

### Model Architecture

A dense model with 27 billion parameters across 64 layers, using a hybrid Gated DeltaNet and Gated Attention design with a 5120 hidden dimension. Features a dual-mode system supporting both thinking mode for complex reasoning and non-thinking mode for efficient dialogue, with a 131K-token context window on Groq.

### Performance Metrics

Qwen 3.6 27B demonstrates flagship-level performance across reasoning and agentic coding benchmarks:

* GPQA Diamond (Reasoning): 87.8%
* AIME 2026 (Math): 94.1%
* LiveCodeBench v6 (Coding): 83.9%
* SWE-bench Verified (Agentic Coding): 77.2%
* SWE-bench Pro (Agentic Coding): 53.5%

### Use Cases

Agentic Coding and Software Engineering

Delivers flagship-level coding performance in a compact dense model, ideal for autonomous coding agents and full-stack development.
* Repository-level code generation and refactoring
* Bug fixing and multi-file edits
* Integration with coding assistants and agent scaffolds
* Tool calling for software engineering workflows

Complex Problem Solving and Dialogue

Switches between deep reasoning and efficient conversation within a single model.
* Multi-step reasoning and mathematical problem solving
* Creative writing and multi-turn dialogue
* Multilingual content generation
* Strategic planning and decision support

Multimodal Visual Understanding

Accepts image and text inputs for vision tasks alongside its text capabilities.
* Image analysis and captioning
* Optical Character Recognition (OCR)
* Visual question answering
* Chart, diagram, and document understanding

### Best Practices

* Mode Selection: use thinking mode (reasoning\_effort="default") for complex reasoning, math, and coding, and non-thinking mode (reasoning\_effort="none") for efficient, general-purpose dialogue.
* Thinking Mode (general): temperature=1.0, top\_p=0.95, top\_k=20, min\_p=0\. For precise coding tasks, lower the temperature to 0.6.
* Non-thinking Mode: temperature=0.7, top\_p=0.80, top\_k=20, min\_p=0, presence\_penalty=1.5.
* Math Problems: include 'Please reason step by step, and put your final answer within \\boxed{}' in the prompt.
* History Management: in multi-turn conversations, only include final outputs without thinking content.
* Reasoning Format: set reasoning\_format to hidden to return only the final answer, or parsed to include the reasoning in a separate field.
* Use the full 131K context window for repository-scale code and multi-document workflows.

### [Get Started with Qwen 3.6 27B](#get-started-with-qwen-36-27b)

Experience state-of-the-art reasoning and agentic coding with Qwen 3.6 27B at Groq speed:

curlJavaScriptPythonJSON

shell

```
pip install groq
```

Python

```
from groq import Groq
client = Groq()
completion = client.chat.completions.create(
    model="qwen/qwen3.6-27b",
    messages=[
        {
            "role": "user",
            "content": "Explain why fast inference is critical for reasoning models"
        }
    ]
)
print(completion.choices[0].message.content)
```