Qwen 3
The multilingual reasoning powerhouse with hybrid thinking modes and 1M+ token context.
Generic Info
- Publisher: Alibaba Cloud
- Release Date: April 2025
- Parameters: 235B Total (22B Active) - Flagship; 0.6B to 235B range
- Context Window: 128K tokens (extendable to 1M+)
- License: Apache 2.0
- Key Capabilities: Hybrid Thinking Modes, 119 Languages, Coding, Math, Agent Tools
Qwen 3 introduces revolutionary hybrid reasoning with "thinking" mode for complex multi-step problems and "non-thinking" mode for fast general responses. The MoE architecture (128 experts, 8 active) delivers frontier performance at a fraction of the compute cost. Support for 119 languages makes it the most multilingual open model available.
Hello World Guide
Run Qwen 3 locally using Hugging Face transformers.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen3-32B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "Solve this step by step: What is 15% of 340?"
messages = [
{"role": "system", "content": "You are a helpful assistant. Think carefully before answering."},
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True # Enable thinking mode for complex tasks
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
**model_inputs,
max_new_tokens=512
)
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
Industry Usage
Complex Problem Solving
Thinking mode excels at multi-step math, scientific reasoning, and strategic planning tasks.
Global Applications
119 language support enables true global deployment for customer service and content generation.
Agentic Workflows
Native tool integration and function calling powers sophisticated AI agents and automation.