3 6 Ke·2025-04-28 23:23Core Insights - Alibaba Cloud has officially open-sourced the Qwen3 series models, which include 2 MoE models and 6 dense models, achieving over 16.9k stars on GitHub within 2 hours of release [2][3] Model Features - The Qwen3 series features 8 parameter sizes ranging from 0.6B to 235B, with flagship models like Qwen3-235B-A22B and Qwen3-30B-A3B showcasing significant capabilities in programming, mathematics, and general reasoning [4][12] - The introduction of a hybrid thinking mode allows users to switch between "thinking" and "non-thinking" modes, enabling control over the depth of reasoning [15][16] - Enhanced reasoning capabilities surpass previous models in mathematics, code generation, and common-sense logic [4][15] Performance Metrics - Qwen3 models have demonstrated superior performance in various benchmarks compared to well-known models such as DeepSeek-R1 and OpenAI's models [12][13] - The Qwen3-30B-A3B model achieves performance exceeding that of QwQ-32B while using only 1/10 of the activated parameters [11][12] - The pre-training dataset for Qwen3 has doubled in size to approximately 3600 billion tokens, enhancing its capabilities in STEM and programming tasks [20][21] Deployment and Accessibility - The Qwen3 models are open-sourced on platforms like Hugging Face, ModelScope, and Kaggle, under the Apache 2.0 license [7] - Developers are encouraged to utilize various frameworks and tools for local deployment, including SGLang and vLLM [9] Future Directions - The company aims to continue enhancing model capabilities by optimizing architecture and training methods, focusing on expanding data scale, increasing model size, and improving long-term reasoning through reinforcement learning [24]