Core Insights - The simultaneous release of important updates by two major AI companies, DeepSeek and Cambricon, indicates a significant collaboration within the Chinese AI industry [5] Group 1: DeepSeek's Developments - DeepSeek officially launched the DeepSeek-V3.2-Exp model on September 29, 2023, which introduces a new sparse attention mechanism aimed at optimizing training and inference efficiency for long texts [1] - The official app, web version, and mini-program have all been updated to the DeepSeek-V3.2-Exp model, and the API costs have been reduced by over 50% [2] Group 2: Cambricon's Response - Cambricon announced that it has achieved compatibility with DeepSeek's latest model, DeepSeek-V3.2-Exp, and has open-sourced the vLLM-MLU inference engine [2] - The company emphasizes its commitment to building a robust software ecosystem for large models and has optimized the performance of DeepSeek models through hardware-software synergy [4] Group 3: Industry Implications - The coordinated release and adaptation actions suggest that leading companies in the Chinese AI industry are engaging in deep collaboration, potentially indicating prior technical discussions between DeepSeek and Cambricon [5]

Core Viewpoint - The simultaneous release of important updates by DeepSeek and Cambricon indicates a deep collaboration among leading companies in the Chinese AI industry, showcasing advancements in model architecture and cost efficiency [7]. Group 1: DeepSeek Updates - DeepSeek officially launched the DeepSeek-V3.2-Exp model, which introduces a Sparse Attention mechanism aimed at optimizing training and inference efficiency for long texts [2]. - The new pricing policy for the DeepSeek API has reduced costs by over 50%, encouraging developers to test the new model [4]. Group 2: Cambricon's Adaptation - Cambricon announced the adaptation of DeepSeek-V3.2-Exp and the open-sourcing of the vLLM-MLU inference engine, allowing developers to experience the new model on Cambricon's hardware platform [4]. - The company emphasized its commitment to software ecosystem development and the synergy between chips and algorithms, which has led to industry-leading computational efficiency levels [6]. Group 3: Industry Implications - The synchronized release and adaptation suggest that Cambricon had been in technical discussions with DeepSeek prior to the model's launch, indicating proactive collaboration in the AI sector [7].

Core Insights - DeepSeek and Cambricon have simultaneously released significant updates regarding their AI models, indicating a strong collaboration within the Chinese AI industry [2][4][6] Group 1: DeepSeek's Model Release - DeepSeek officially launched the DeepSeek-V3.2-Exp model on September 29, 2025, which introduces a new sparse attention mechanism aimed at improving training and inference efficiency for long texts [2][3] - The new model version is available across DeepSeek's app, web, and mini-program platforms, with a significant reduction in API costs by over 50% for developers [3][6] Group 2: Cambricon's Adaptation and Support - Cambricon announced the immediate adaptation of DeepSeek-V3.2-Exp and has open-sourced the vLLM-MLU inference engine, allowing developers to experience the new model on Cambricon's hardware platform [4][5] - The company emphasizes its commitment to building a robust software ecosystem for large models, supporting mainstream open-source models like DeepSeek [5][6] Group 3: Technical Innovations and Collaboration - Cambricon has focused on the joint innovation of chips and algorithms to optimize large model deployment performance and reduce costs, achieving industry-leading computational efficiency levels [6] - The collaboration between DeepSeek and Cambricon suggests that technical discussions and adaptation efforts may have begun prior to the official release of DeepSeek-V3.2 [6]

9月29日寒武纪-U大宗交易一览 | 成交量 | 成交金额 | 成交价格 | 相对当日收盘折 | 买方营业部 | 卖方营业部 | | --- | --- | --- | --- | --- | --- | | （万股） | （万元） | （元） | 溢价（%） | | | | 1.52 | 2011.72 | 1323.50 | 0.00 | 国泰海通证券股份有 | 中信证券股份有限公 | | | | | | 限公司总部 | 司上海分公司 | 寒武纪-U9月29日大宗交易平台出现一笔成交，成交量1.52万股，成交金额2011.72万元，大宗交易成交 价为1323.50元。该笔交易的买方营业部为国泰海通证券股份有限公司总部，卖方营业部为中信证券股 份有限公司上海分公司。 （文章来源：证券时报网） 进一步统计，近3个月内该股累计发生12笔大宗交易，合计成交金额为1.46亿元。 证券时报·数据宝统计显示，寒武纪-U今日收盘价为1323.50元，下跌1.16%，日换手率为2.16%，成交额 为118.16亿元，全天主力资金净流入6161.91万元，近5日该股累计下跌5.33%，近5日资金合计净流入 4.66亿元。 两 ...

Group 1 - The CNQQ ETF, focused on Chinese technology, was launched on September 26 on NASDAQ, aiming to provide global investors with exposure to China's tech and innovation sectors [1] - The underlying index, Solactive ChinaAMC Transformative China Tech Index, was developed in collaboration with Solactive AG and China Asset Management, emphasizing companies with strong R&D capabilities [2] - The index uses a non-traditional market capitalization weighting method, selecting the top 100 stocks based on adjusted market cap and R&D spending, with a maximum weight of 10% per stock [2] Group 2 - The Solactive ChinaAMC Transformative China Tech Index includes nearly 100 Chinese companies listed in mainland China and Hong Kong, spanning five sectors: automotive and transportation, commercial and consumer services technology, electronic and electrical products, healthcare technology, and industrial and manufacturing technology [2] - Major holdings in the CNQQ ETF include Alibaba Group (10.94%), Tencent Holdings (9.93%), and Contemporary Amperex Technology (8.00%) [4] Group 3 - Morgan Stanley noted a shift in investor sentiment towards Chinese technology since the "9·24" event, indicating a cautious optimism regarding the Chinese stock market and improving corporate earnings in various sectors [5] - The Hong Kong technology fund has seen significant inflows, ranking first in capital inflow among single market sector funds, while U.S. technology funds have experienced outflows [8]

Core Insights - DeepSeek has officially released and open-sourced the DeepSeek-V3.2-Exp model on the Hugging Face platform, marking a significant step towards the next generation architecture [1] - The new model introduces the DeepSeek Sparse Attention (DSA) mechanism, which aims to optimize training and inference efficiency for long texts while reducing computational resource consumption [1] - The model supports a maximum context length of 160K, with successful adaptations completed by Huawei and Cambricon [1] Technical Breakthroughs - The DeepSeek Sparse Attention (DSA) mechanism achieves fine-grained sparse attention, significantly enhancing training and inference efficiency for long text scenarios without compromising output quality [1][3] - The training settings for DeepSeek-V3.2-Exp were strictly aligned with the previous version, V3.1-Terminus, showing comparable performance across major public evaluation datasets [3] Benchmark Performance - Performance comparison between DeepSeek-V3.1-Terminus and DeepSeek-V3.2-Exp across various benchmarks shows: - MMLU-Pro: 85.0 (both versions) - GPQA-Diamond: 80.7 (V3.1) vs 79.9 (V3.2) - Humanity's Last Exam: 21.7 (V3.1) vs 19.8 (V3.2) - BrowseComp: 38.5 (V3.1) vs 40.1 (V3.2) - SimpleQA: 96.8 (V3.1) vs 97.1 (V3.2) - Codeforces-Div1: 2046 (V3.1) vs 2121 (V3.2) - AIME 2025: 88.4 (V3.1) vs 89.3 (V3.2) [4] Cost Reduction - The introduction of the new model has led to a significant reduction in API service costs, with a price drop of over 50%, effective immediately [4] Open Source and Community Support - DeepSeek has fully open-sourced the DeepSeek-V3.2-Exp model on Hugging Face and ModelScope, along with related research papers [6] - The company has retained API access for the V3.1-Terminus version for comparison purposes until October 15, 2025, with pricing aligned to V3.2-Exp [6] - To support community research, DeepSeek has also open-sourced GPU operators designed for the new model, recommending the use of the TileLang version for ease of debugging and rapid iteration [6] Industry Collaboration - Cambricon has announced the completion of adaptation for the new model and has open-sourced the vLLM-MLU inference engine source code, allowing developers to experience the new model's features on their hardware platform [6][7]

Core Insights - DeepSeek has officially released the DeepSeek-V3.2-Exp model, which introduces a Sparse Attention mechanism for improved training and inference efficiency on long texts [1] - The official app, web version, and mini-program have all been updated to DeepSeek-V3.2-Exp, with a significant reduction in API costs by over 50% for developers [1] - The company has also adapted to the latest DeepSeek model and open-sourced the vLLM-MLU inference engine code, allowing developers to experience the new model on their platform [1] Model Iteration and Features - DeepSeek is progressing with model iterations, having recently upgraded to DeepSeek-V3.1-Terminus, which features a hybrid inference architecture supporting both thinking and non-thinking modes [2] - The V3.1 model boasts higher thinking efficiency and enhanced agent capabilities, showing significant improvements in tool usage and intelligent task performance [2] - The V3.1 model utilizes UE8M0 FP8 Scale parameter precision, designed for the upcoming generation of domestic chips, which has positively impacted the stock prices of related domestic chip industry companies [2]

Core Insights - DeepSeek has released the experimental version DeepSeek-V3.2-Exp, which introduces Sparse Attention for improved training and inference efficiency on long texts [1] - The API pricing has been reduced by over 50% due to significant cost savings from the new model [1] - Cambricon has adapted to DeepSeek-V3.2-Exp and open-sourced the vLLM-MLU inference engine, allowing developers to experience the new model on their platform [1][2] - Huawei Ascend has also quickly adapted to DeepSeek-V3.2-Exp, open-sourcing all inference code and achieving optimized deployment on the CANN platform [3] Group 1 - DeepSeek-V3.2-Exp is an experimental version that builds on the previous V3.1-Terminus, focusing on optimizing long text processing [1] - The new model's API pricing reduction is a strategic move to enhance developer engagement and usage [1] - Cambricon's rapid adaptation to the new model demonstrates its commitment to software ecosystem development and performance optimization [2] Group 2 - Huawei's deployment of DeepSeek-V3.2-Exp BF16 model showcases its capability in handling large sequence processing with low latency and high throughput [3] - The continuous iteration of DeepSeek models indicates a proactive approach to addressing user feedback and improving model performance [3]

Core Insights - DeepSeek Company released its new model architecture DeepSeek-V3.2, which has garnered significant industry attention [1] - Cambricon announced its adaptation to DeepSeek-V3.2 and open-sourced the large model inference engine vLLM [1][2] - The DeepSeek-V3.2-Exp model introduces DeepSeek Sparse Attention, optimizing training and inference efficiency for long texts [1] Company Developments - DeepSeek-V3.2-Exp is an experimental version built on V3.1-Terminus, focusing on sparse attention mechanisms [1] - The official DeepSeek applications and APIs have been updated to V3.2-Exp, with significant price reductions to encourage user testing and feedback [1] - Cambricon's adaptation to DeepSeek-V3.2-Exp indicates prior collaboration and technical communication between the two companies [2] Industry Trends - The rapid adaptation of Cambricon to the new model reflects a significant signal of deep collaboration among top Chinese tech companies [2] - The large model has a substantial size of 671GB, requiring approximately 8-10 hours to download under ideal bandwidth conditions [2] - The collaboration between leading companies in the AI chip and model sectors is seen as a strong example of innovation and cooperation in China's tech industry [2]

Core Viewpoint - The release of DeepSeek-V3.2 by DeepSeek Company and its adaptation by Cambricon signifies a strong collaboration among leading tech firms in China's AI industry, aiming to enhance efficiency in long-text training and inference [2][3][4]. Group 1: Model Release and Features - DeepSeek Company launched the experimental version DeepSeek-V3.2-Exp, which introduces a sparse attention mechanism for optimizing long text training and inference [2]. - The new model has a substantial size of 671GB, requiring approximately 8-10 hours for download under ideal bandwidth conditions [3]. Group 2: Collaboration and Industry Impact - Cambricon's quick adaptation to DeepSeek-V3.2-Exp indicates prior collaboration and communication between the two companies, reflecting a trend of low-profile yet effective partnerships in the tech industry [3]. - The collaboration between leading companies in the AI model and chip sectors is expected to significantly reduce training and inference costs for users, facilitating the emergence of AI applications [4].