Groq
Login
Loading model information...

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

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

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)

Was this page helpful?