Qwen3.6-Plus
Qwen3 is a family of large language models from Alibaba's Qwen team, including dense and Mixture-of-Experts (MoE) models in various sizes. It supports both thinking and non-thinking modes, enabling complex reasoning and efficient general chat. The models are designed for developers and researchers who need state-of-the-art open-weight LLMs for a wide range of tasks.
✨ Key features
- Dense and MoE models from 0.6B to 235B-A22B
- Seamless switching between thinking and non-thinking modes
- Enhanced reasoning, math, coding, and tool usage
- Support for 100+ languages and dialects
- Long-context understanding up to 1 million tokens
- Superior human preference alignment for creative tasks
🎯 Use cases
- Complex logical reasoning and mathematical problem solving
- Code generation and software development assistance
- Multilingual translation and instruction following
- Agent-based tasks with external tool integration
- Creative writing, role-playing, and multi-turn dialogue
📦 Installation
🧰 Requirements: Python with Transformers >=4.51.0, or llama.cpp >=b5401, or Ollama v0.9.0+; GPU recommended for larger models.
# Install Transformers
pip install transformers>=4.51.0
# For llama.cpp, build from source or download a release
# For Ollama, install from https://ollama.com/
🚀 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen3-30B-A3B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
prompt = "Give me a short introduction to large language model."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(**model_inputs, max_new_tokens=16384)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
content = tokenizer.decode(output_ids, skip_special_tokens=True)
print(content)
❓ FAQ
What is the difference between Qwen3-Instruct-2507 and Qwen3-Thinking-2507?
Instruct-2507 is the non-thinking mode optimized for general chat and instruction following, while Thinking-2507 is the thinking mode that generates reasoning before answering, excelling in complex tasks.
How do I enable or disable thinking mode?
For previous Qwen3 models, you can pass enable_thinking=False to apply_chat_template or use /think and /no_think in the prompt. For 2507 versions, the mode is fixed per model.
What are the hardware requirements?
The README does not specify exact hardware, but larger models like 235B-A22B require substantial GPU memory. Refer to the speed benchmark documentation for details.
Can I use Qwen3 with llama.cpp or Ollama?
Yes, llama.cpp (>=b5401) and Ollama (v0.9.0+) are supported. For llama.cpp, use the -hf flag to load GGUF models; for Ollama, use ollama run qwen3:8b.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.