Tai Mei Ti A P P·2025-04-28 23:27Core Insights - Alibaba has launched its new open-source model Qwen3, which is claimed to be the strongest open-source model globally, outperforming competitors like DeepSeek-R1 and OpenAI-o1 with a total parameter count of 235 billion and activation of only 22 billion [1][3]. Model Performance - Qwen3 includes eight models, with two MoE models of 30B and 235B parameters, and six dense models ranging from 0.6B to 32B, achieving state-of-the-art performance in their respective sizes [3]. - The 30B MoE model of Qwen3 shows over ten times performance leverage, requiring only 3B activation to match the performance of the previous Qwen2.5-32B model [3]. - In various assessments, Qwen3 scored 81.5 in the AIME25 evaluation and surpassed 70 in the LiveCodeBench, outperforming Grok3 [5]. Training Data and Methodology - Qwen3 was pre-trained on approximately 36 trillion tokens, nearly double the amount used for Qwen2.5, enhancing its language skills and general knowledge [8][9]. - The training process involved three phases, focusing on diverse long reasoning chain data, large-scale reinforcement learning, and integrating thinking modes for seamless reasoning and quick responses [11][13]. Deployment and Cost Efficiency - The deployment cost for Qwen3 is significantly reduced, requiring only four H20 GPUs for full deployment, which is one-third of the memory usage compared to similar performance models [5]. - Qwen3 supports 119 languages and dialects and is available under the Apache 2.0 license on platforms like Hugging Face and ModelScope [7]. Agent Capabilities - Qwen3 is the first hybrid reasoning model in China, integrating "fast thinking" and "slow thinking" modes, allowing for quick responses to simple queries and deep reasoning for complex problems [4]. - The model's agent capabilities were evaluated with a score of 70.8 in the BFCL assessment, surpassing top models like Gemini2.5-Pro and OpenAI-o1 [7].