Qwen (Tongyi Qianwen)

Imagine Alibaba, one of the biggest tech companies in the world, deciding to open-source its most advanced AI models completely for free. Their absolute best stuff, from tiny 0.5B parameter models that run on your phone all the way up to massive 235B MoE beasts that compete with the best models money can buy. That's Qwen.

Qwen (also known as Tongyi Qianwen, meaning "Unified Thousand Questions") is a family of LLMs developed by Alibaba Cloud. First announced in April 2023, Qwen has rapidly evolved through multiple major versions, becoming one of the most popular open-weight model families in the world.

Qwen covers dense models from 0.5B to 72B, MoE models up to 235B-A22B, vision-language models, audio models, coding-specialized models, and even "Omni" models handling text, images, video, and audio. All released under the permissive Apache 2.0 license (starting with Qwen3). By mid-2026, Qwen models had been downloaded hundreds of millions of times on Hugging Face.

In April 2023, Alibaba Cloud announced Tongyi Qianwen, its LLM designed to compete with Baidu, Tencent, and global foundation models. The initial beta launched through Alibaba's DingTalk platform and quickly gained traction among Alibaba's vast enterprise ecosystem.

The first open-source model, Qwen-7B, was released in August 2023. Qwen1.5 (Feb 2024) offered 9 sizes from 0.5B to 72B. Qwen2 (Jun 2024) introduced Grouped Query Attention and 128K context. Qwen2.5 (Sep 2024) brought major improvements in knowledge, coding, and math, along with specialized variants for code, math, vision, and Omni.

Qwen3 (Apr 2025) was a generational leap introducing thinking + non-thinking modes, released under Apache 2.0. The Qwen3-2507 update (Jul-Aug 2025) brought 1M token context and SOTA results among open-weight thinking models.

Model	Total Params	Active Params	Type
Qwen3-0.6B	0.6B	0.6B	Dense
Qwen3-8B	8B	8B	Dense
Qwen3-32B	32B	32B	Dense
Qwen3-30B-A3B	30B	3B	MoE
Qwen3-235B-A22B	235B	22B	MoE

The real innovation in Qwen3 is seamless switching between thinking and non-thinking modes. In thinking mode, the model generates internal reasoning traces inside <think> tags. In non-thinking mode, it responds directly. Switch per-query with /think or /no_think.

• Dense and MoE Architectures - Both supported with standard transformers and sparsely-gated MoE carefully tuned for load balancing.
• Thinking/Non-Thinking Modes - Dynamic switching between reasoning and direct response per query. Unique to Qwen3.
• Multimodal (Qwen2.5-Omni) - Single model handling text, images, video, and audio inputs with text and audio outputs.
• Long Context - Progressive extension from 8K to 1M tokens across model generations.
• Apache 2.0 License - Starting with Qwen3, one of the most permissive open-source licenses available.

Over 90,000 corporate clients use Qwen through Alibaba's Model Studio platform. Over 2.2 million corporate users access Qwen through DingTalk. Applications span consumer electronics, automotive in-car assistants, online gaming NPC dialogs, enterprise document processing, and e-commerce.

Qwen supports over 100 languages and dialects. The training data includes massive amounts of multilingual text, making Qwen one of the most language-diverse open-weight model families. Particularly strong in Asian languages (Chinese, Japanese, Korean, Vietnamese, Thai) and European languages (English, Spanish, French, German, Arabic).

Qwen3 introduced advanced agent capabilities with precise tool integration in both thinking and non-thinking modes. The Qwen-Agent framework at github.com/QwenLM/Qwen-Agent supports tool use, code execution, and multi-agent collaboration.

Qwen models are available on Hugging Face and ModelScope. Supported by vLLM, SGLang, TGI, llama.cpp, Ollama, LM Studio, MLX, and OpenVINO. Fine-tuning frameworks include Axolotl and LLaMA-Factory. Active community on Discord and GitHub.

vs DeepSeek - Qwen's advantage is broader model range, superior enterprise support through Alibaba Cloud, and thinking/non-thinking switching. DeepSeek pushes the efficiency frontier.

vs Llama (Meta) - Qwen excels in multilingual capabilities (especially Asian languages), offers more model sizes, and has thinking/non-thinking mode innovation that Llama lacks. On benchmarks they trade blows.

vs GPT-4o and Claude - Frontier models still lead on some creative tasks and safety alignment. But Qwen3-235B-A22B competes closely on reasoning, math, and coding benchmarks, and is fully open-weight with Apache 2.0.

Quick Start with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-8B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Explain the transformer architecture in simple terms."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Run with Ollama:

ollama run qwen3:8b

Try Qwen for free at chat.qwen.ai.