Late Night Raid, Alibaba Qwen3 Tops the Global Open Source Throne!#
Just early this morning, the highly anticipated new generation of Alibaba's Tongyi Qianwen model, Qwen3, was open-sourced! Upon its release, it immediately ascended to the throne of the world's strongest open-source model. Its parameter count is only 1/3 of DeepSeek-R1, but the cost has significantly decreased, and its performance has comprehensively surpassed that of R1, OpenAI-o1, and other top global models.
Features of Qwen3#
- Hybrid Reasoning Model: The first "hybrid reasoning model" in China, integrating "fast thinking" and "slow thinking" into a single model.
- For simple needs, it can provide answers with low computational power in "seconds."
- For complex problems, it can perform multi-step "deep thinking," greatly saving computational resources.
- Architecture and Parameters: Utilizes a mixture of experts (MoE) architecture, with a total parameter count of 235B, activating only 22B.
- Pre-training Data: The pre-training data volume reaches 36T, and in the post-training phase, multiple rounds of reinforcement learning seamlessly integrate non-thinking modes into the thinking model.
Performance#
- Benchmark Testing: Qwen3, upon its birth, immediately swept through major benchmarks.
- Deployment Cost: While performance has significantly improved, deployment costs have drastically decreased, requiring only 4 H20 to deploy the full version of Qwen3, with memory usage being only 1/3 of similar performance models!
Highlights Summary#
- Various sizes of dense models and mixture of experts (MoE) models, including 0.6B, 1.7B, 4B, 8B, 14B, 32B, as well as 30B-A3B and 235B-A22B.
- Capable of seamless switching between thinking mode (for complex logical reasoning, mathematics, and coding) and non-thinking mode (for efficient general chatting).
- Significantly enhanced reasoning capabilities, surpassing the previous QwQ in thinking mode and the Qwen2.5 instruct model in non-thinking mode.
- More aligned with human preferences, excelling in creative writing, role-playing, multi-turn dialogue, and instruction following.
- Proficient in AI agent capabilities, supporting precise integration with external tools in both thinking and non-thinking modes.
- First to support 119 languages and dialects, with strong multilingual instruction following and translation capabilities.
Access Links#
- Online Experience: Qwen3 Online Experience
- Magic Community: Magic Community
- Hugging Face: Hugging Face
- GitHub: GitHub
As of now, Alibaba Tongyi has open-sourced over 200 models, with global downloads exceeding 300 million, and the number of Qianwen derivative models surpassing 100,000, completely surpassing the American Llama, becoming the world's first open-source model!
Qwen3 Family Debuts#
This time, Alibaba has open-sourced 8 hybrid reasoning models at once, including 2 MoE models of 30B and 235B, as well as 6 dense models of 0.6B, 1.7B, 4B, 8B, 14B, and 32B, all under the Apache 2.0 license.
- The 30B parameter MoE model of Qwen3 achieves over 10 times the model performance leverage, with only 3B activated, comparable to the performance of the previous Qwen2.5-32B model.
- The performance of Qwen3's dense models continues to break through, achieving the same high performance with half the parameter count.
New King of Open Source Models, Breaking Records#
Qwen3 has significantly enhanced reasoning, instruction following, tool calling, and multilingual capabilities, setting a new performance high for all domestic models and global open-source models—scoring 81.5 points in the AIME25 evaluation at the level of Olympiad mathematics, breaking the open-source record.
- The flagship model Qwen3-235B-A22B has shown impressive results in coding, mathematics, and general capabilities across various benchmark tests compared to other top models (such as DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro).
Main Features#
Hybrid Reasoning Mode#
The Qwen3 model introduces a hybrid problem-solving approach, supporting two modes:
- Thinking Mode: Step-by-step reasoning, suitable for complex problems.
- Non-Thinking Mode: Quickly provides answers, suitable for simple problems.
This flexibility allows users to control the model's reasoning process based on the complexity of the task.
Multilingual Support#
The Qwen3 model supports 119 languages and dialects, with great potential for creating international applications.
More Powerful Agent Capabilities#
Alibaba has optimized the Qwen3 model to enhance coding and agent capabilities, and has strengthened support for MCP.
Rave Reviews Across the Internet#
Within less than 3 hours of its open-source release, Qwen3 garnered 17k stars on GitHub, igniting enthusiasm in the open-source community. Developers are downloading it in droves, starting rapid testing.
- Project Address: GitHub