Core Insights - The release of the Step 3.5 Flash model by Jieyue Xingchen has rapidly gained popularity, achieving the fastest model ranking on OpenRouter and topping the global trending list within two days [1][3][28] Model Performance - Step 3.5 Flash has demonstrated impressive performance in mathematical reasoning with a score of 97.3 on AIME 2025 and a 74.4% success rate in code repair [4] - The model utilizes a sparse mixture of experts (MoE) architecture, allowing it to dynamically select the most suitable experts for each token while maintaining a total parameter count of 196 billion [4][5] - It supports a high throughput of 100-300 tokens per second (TPS), with some scenarios achieving up to 350 TPS, significantly surpassing the previous year's range of 50-100 TPS [7][9] Technical Innovations - The model incorporates MTP-3 technology, enabling it to predict multiple tokens simultaneously, enhancing both speed and logical coherence during multi-turn interactions [9][26] - Step 3.5 Flash effectively handles long contexts through a hybrid architecture of sliding windows and global attention, allowing for efficient processing of up to 256K tokens [7][26] User Adoption and Feedback - Developers and users prioritize practical performance over benchmark scores, as evidenced by the widespread adoption of Step 3.5 Flash in various AI applications and workflows [21][28] - The model's ability to generate reliable outputs and perform well in multilingual contexts has been positively noted by users, further solidifying its reputation [14][21] Industry Trends - The shift in focus from conversational AI to more complex task handling reflects the evolving needs of developers and users, emphasizing the importance of efficiency and reliability in AI applications [27][29] - The success of Step 3.5 Flash indicates a broader industry trend towards models that prioritize practical utility over mere technical prowess [29]
全球AI开发者新宠：阶跃星辰Step 3.5 Flash，两天登顶OpenRouter趋势榜
3 6 Ke·2026-02-07 05:05