实测！Qwen下一代基础架构突袭！秒解AIME数学竞赛题，提速10倍+性价比提升10倍

Core Insights - The article discusses the release of Qwen3-Next, a next-generation model architecture, which is a preview of Qwen3.5 [1] - The Qwen team has open-sourced the Qwen3-Next-80B-A3B-Base model, which has 80 billion parameters but costs less than one-tenth of the training cost of Qwen3-32B [2][3] - The new model demonstrates significant improvements in context processing and inference efficiency, achieving up to ten times the throughput of its predecessor in long context scenarios [3][24] Model Improvements - Hybrid Attention Mechanism: The Qwen3-Next model incorporates a Gated DeltaNet for better context learning, using a 3:1 hybrid strategy to balance performance and efficiency [10] - High Sparsity MoE Structure: The model features a high sparsity MoE architecture with 80 billion total parameters, activating only about 3 billion during inference [13] - Stability Optimization: The model employs Zero-Centered RMSNorm and weight decay to enhance training stability and mitigate weight growth issues [16][17] - Multi-Token Prediction Mechanism: The introduction of a native Multi-Token Prediction mechanism improves overall model performance and speculative decoding acceptance rates [18] Performance Metrics - Qwen3-Next-80B-A3B-Base outperforms Qwen3-32B-Base in most benchmark tests while using only 9.3% of the GPU resources required by Qwen3-32B [22][28] - In various benchmarks, Qwen3-Next-80B-A3B-Instruct shows superior performance compared to Qwen3-30B-A3B-Instruct-2507 and approaches the performance of Qwen3-235B-A22B-Instruct-2507 [31][34] - Qwen3-Next-80B-A3B-Thinking surpasses the closed-source model Gemini-2.5-Flash-Thinking in multiple benchmark tests [35] Practical Applications - The model supports multimodal capabilities, allowing for quick and accurate responses to complex tasks, such as solving math problems and generating code [39][43] - Users can access the new model through various platforms, including Qwen Chat and API services provided by Alibaba Cloud [48]