Market Performance Review - A-shares continued to rise strongly, reaching new highs, with major indices all increasing, particularly the Sci-Tech 50 which surged by 8.59% on Friday and 13.31% for the week, significantly outperforming other indices [3][12] - The market style saw a significant reversal, with cyclical stocks rebounding strongly while consumer stocks lagged behind, contrasting with the previous week [12] - Large-cap stocks outperformed small-cap stocks this week, reversing last week's trend, with core assets like the Moutai index and the Ning combination also seeing substantial gains, up 4.51% and 3.71% respectively [12] Industry Insights - The TMT sector continued to lead the market, with significant gains in the communication (10.84%), electronics (8.95%), and computer (7.93%) sectors, driven by the production halt of Nvidia's H20 chips and optimistic expectations for domestic AI applications and computing power [4][13] - The expectation of a Federal Reserve interest rate cut is anticipated to further boost A-shares, influencing both short-term capital flows and long-term fundamental changes [4][30] Future Outlook and Investment Views - The expectation of Federal Reserve rate cuts is likely to enhance A-share performance through improved capital flows and a more attractive investment environment for international capital [4][30] - Long-term fundamental changes include a potential recovery in consumer and investment demand globally, alleviating export pressures, and a possible domestic rate cut that could reverse current deleveraging trends [4][31] - The report emphasizes that individual stock alpha logic is preferred over industry beta logic, highlighting opportunities for valuation recovery in TMT growth sectors, particularly in AI applications and computing power [5][31]

Report Industry Investment Rating - Not provided in the content Core Viewpoints of the Report - After two consecutive weeks of adjustment in the bond market since mid - August, the decline has exceeded the previous round in late July, resulting in a certain degree of cost - effectiveness. Currently, the strategy should prioritize liquidity. There are opportunities in 3 - 5 - year bank secondary capital bonds after adjustment, and it is also advisable to participate in the sinking of weak - quality urban investment bonds with a maturity of 1 - 3 years. However, the ultra - long - term strategy may not be a good choice due to high market uncertainty [3][36] Summary by Relevant Catalogs 1. Market Adjustment and Bond Performance - Since mid - August, the bond market has been continuously adjusting for two weeks, especially last week's adjustment exceeding expectations. Credit bonds declined synchronously, and the decline of major maturity varieties was higher than that of interest rates. The stock - bond "seesaw" effect continued, with the Shanghai Composite Index hitting a new high, and the bond market being insensitive to fundamental indicators, resulting in a continuous decline and rising yields [1][9] - From August 18 to 22, 2025, the yields of 1Y, 2Y, 3Y, 4Y, and 5Y treasury bonds increased by 0.4BP, 3.2BP, 9.7BP, 8.1BP, and 3.8BP respectively. The yields of AAA medium - and short - term notes with the same maturities increased by 4.9BP, 6.6BP, 5.8BP, 7.6BP, and 4.6BP respectively, and the yields of AA+ medium - and short - term notes increased by 4.9BP, 6.6BP, 7.8BP, 6.6BP, and 5.6BP respectively [9][10] - The market of ultra - long - term credit bonds weakened synchronously, with most of the declines exceeding those of the same - maturity interest - rate bonds. The decline of highly liquid ultra - long - term secondary and perpetual bonds was the lowest, while the decline of ultra - long - term urban investment bonds with the poorest liquidity was relatively large. The yields of AAA/AA+ 10Y medium - term notes increased by 6.00BP and 7.00BP respectively, and the yields of AAA/AA+ 10Y urban investment bonds increased by 13.01BP and 11.00BP respectively. The yield of AAA - 10Y bank secondary capital bonds increased by 6.69BP, while the yield of 10Y treasury bonds increased by 3.53BP [11][12] 2. Performance of Secondary and Perpetual Bonds - The market of secondary and perpetual bonds weakened synchronously, but the "volatility amplifier" feature was not obvious. The declines of 1Y - 5Y were similar to those of general credit bonds, and the decline gap in the ultra - long - term part was also close to that of ultra - long - term credit bonds. Currently, the part of the curve with a maturity of 3 years and above is still 25BP - 35BP away from the lowest yield point since 2025. Compared with the sharp decline at the end of July, the yield points of bonds with a maturity of over 3 years have reached new highs, and the adjustment amplitude is higher than that of the sharp decline at the end of July [2][16] - In terms of active trading, the sentiment was the most pessimistic in the second week of August. Although the market was still adjusting last week, the marginal sentiment of secondary and perpetual bonds improved. From August 11 to 15, the proportion of low - valuation transactions of secondary and perpetual bonds was 5.00%, 0.00%, 100.00%, 5.00%, and 0.00% respectively, and the average trading duration was 0.74 years, 1.02 years, 3.81 years, 1.53 years, and 1.12 years respectively. From August 18 to 22, the proportion of low - valuation transactions was 0.00%, 100.00%, 17.07%, 100.00%, and 100.00% respectively, and the average trading duration was 0.65 years, 4.73 years, 1.03 years, 5.66 years, and 3.30 years respectively [2][18] 3. Institutional Behavior - Public funds and other trading desks continued to sell, but it was more of a portfolio rebalancing rather than a full - scale reduction. At the same time, allocation desks such as wealth management and insurance institutions moderately bought during the adjustment. Public funds reduced their holdings of secondary bonds of national and joint - stock banks with a maturity of 3 - 5 years, with the total selling scale in the past two weeks approaching 20 billion, but they also increased their holdings of secondary capital bonds with a maturity of 1 - 3 years. Public funds were not very willing to sell their core assets such as weak - quality urban investment bonds [3][29] - Allocation desks such as bank wealth management and insurance institutions bought opportunistically after the sharp decline in the bond market, but they were also cautious about the maturity, mainly focusing on varieties with a maturity of 3 years and below. Since August, the increase in the liability side of wealth management products has been limited, and the demand is not strong, but it is not a full - scale redemption [3][29] 4. Performance of Credit Bond ETF Products - Credit bond ETF products performed poorly during the market adjustment in the past two weeks, with weak scale growth and net - value performance. In terms of scale change, the weekly scale of credit benchmark market - making ETF products has shrunk for two consecutive weeks since the market adjustment in the second week of August, and the weekly scale of science and technology innovation ETF products has been significantly weaker in August than in July. In terms of unit net - value change, the unit net values of the above two types of credit bond ETFs have suffered losses for two consecutive weeks, and the loss scale increased last week. In addition, the average turnover rate of the above two types of credit bond ETFs dropped to a new low last week [33]

个股表现 2024-08 2024-11 2025-01 2025-03 2025-06 2025-08 -6% 11% 28% 45% 62% 79% 96% 113% 130% 147% 工业富联 电子 资料来源：聚源，中邮证券研究所 公司基本情况 | 最新收盘价（元） | 48.00 | | --- | --- | | 总股本/流通股本（亿股）198.59 | / 198.58 | | 总市值/流通市值（亿元）9,533 | / 9,532 | | 52 周内最高/最低价 | 48.91 / 15.44 | | 资产负债率(%) | 51.8% | | 市盈率 | 41.03 | | 第一大股东 | 富泰华工业(深圳)有限 | | 公司 | | 证券研究报告：电子 | 公司点评报告 发布时间：2025-08-26 股票投资评级 买入 |维持 研究所 分析师:吴文吉 SAC 登记编号:S1340523050004 Email:wuwenji@cnpsec.com 工业富联(601138) AI 服务器需求强劲，GB200 系列良率持续改善 l 事件 公司发布 2025 年半年报，上半年实现营业收入 360 ...

发布时间：2025-08-26 研究所 分析师:梁伟超 SAC 登记编号:S1340523070001 Email:liangweichao@cnpsec.com 研究助理:谢鹏 SAC 登记编号:S1340124010004 Email:xiepeng@cnpsec.com 近期研究报告 《风险偏好如何定价？——流动性周 报 20250824》 - 2025.08.25 证券研究报告：固定收益报告 固收周报 长久期中低评级产业债流动性下降 ——流动性打分周报 20250825 ⚫ 核心解读 本周报以 qb 的债券资产流动性打分为基础，跟踪不同债券板块 个券的流动性得分情况。 城投债方面，分区域看，江苏高等级流动性债项数量有所增加， 四川、天津、重庆整体维持，山东有所减少。分期限看，1 年以内、 2-3 年期高等级流动性债项数量有所增加，1-2 年期整体维持，3-5 年 和 5 年期以上有所减少。从隐含评级看，隐含评级为 AA(2)的高等级 流动性债项数量有所增加，隐含评级为 AA+的高等级流动性债项数量 整体维持，隐含评级为 AAA、AA、AA-的高等级流动性债项数量有所减 少。 产业债方面，分行业看，公用 ...

证券研究报告：农林牧渔 | 公司点评报告 股票投资评级 公司基本情况 分析师:王琦 SAC 登记编号:S1340522100001 Email:wangqi2022@cnpsec.com 牧原股份(002714) 养殖成本优势突出，高分红积极回报股东 事件： 公司发布 25 年中报，实现营收 764.63 亿元，同比增 34.46%；归 母净利 105.30 亿元，同比增 1169.77%，处于业绩预告中上区间。成 本快速下行，推动公司业绩大增。同时公司财务状况良好，上半年经 营活动产生的净现金流为 173.51 亿元，同比增 12.13%；截至二季度 末，公司资产负债率为 56.06%，相比一季度末下降 3.14 个百分点。 点评：养殖成本优势突出，屠宰业务大幅减亏 养殖：出栏稳增，成本优势铸造最强护城河。2025 年上半年， 公司共销售生猪 4691 万头(YOY+44.84%)，其中商品猪 3839.4 万头 (YOY+32.48%)，仔猪 829.1 万头(YOY+168.06%)。公司生产成绩持续 改善，养殖成本从 1 月的 13.1 元/公斤逐月降至 7 月的 11.8 元/公 斤，稳居行业第一梯 ...

Investment Rating - The investment rating for the company is "Buy" [13] Core Views - The company reported a revenue of 10.55 billion yuan for the first half of 2025, a year-on-year increase of 2.17%, but the net profit attributable to shareholders decreased by 32.69% to 288 million yuan [5][6] - The decline in net profit was primarily due to a reduction in government subsidies, which decreased by 144 million yuan year-on-year [6] - New orders remained stable, with a total of 14.38 billion yuan in new orders signed in the first half of 2025, a slight increase of 0.2% year-on-year [6] - The company has made progress in robotics, having developed a welding robot control system and begun external sales [7] Financial Summary - The company’s total market capitalization is 12.7 billion yuan, with a total share capital of 690 million shares [4] - The company’s debt-to-asset ratio stands at 61.9% [4] - The projected revenue for 2025 and 2026 is expected to be 22.2 billion yuan and 23 billion yuan, respectively, with growth rates of 3.4% and 3.5% [7][9] - The estimated net profit for 2025 is 680 million yuan, reflecting a decrease of 12.1%, while the profit for 2026 is projected to increase by 62% to 1.1 billion yuan [7][9]

Quantitative Models and Construction 1. Model Name: GRU Model - **Model Construction Idea**: The GRU model is used to predict stock returns based on historical data and incorporates various factors to optimize portfolio performance [3][4][5] - **Model Construction Process**: - The GRU model is trained on historical data to capture temporal dependencies in stock returns - It uses multiple input features, including technical and fundamental factors, to predict future returns - The model is applied to different stock pools (e.g., CSI 300, CSI 500, CSI 1000) to evaluate its performance [5][6][7] - **Model Evaluation**: The GRU model demonstrates strong performance in most stock pools, with positive long-short returns across various factors. However, certain sub-models (e.g., `barra5d`) show occasional underperformance [5][6][7] 2. Model Name: Open1d and Close1d Models - **Model Construction Idea**: These models focus on short-term price movements and are designed to capture daily return patterns [8][31] - **Model Construction Process**: - Open1d and Close1d models are trained on daily open and close price data, respectively - They are evaluated based on their ability to generate excess returns relative to the CSI 1000 index [8][31] - **Model Evaluation**: These models show mixed performance, with occasional drawdowns relative to the benchmark index [8][31] 3. Model Name: Barra1d and Barra5d Models - **Model Construction Idea**: These models are based on the Barra factor framework and aim to capture short-term and medium-term return patterns [8][31] - **Model Construction Process**: - Barra1d focuses on daily factor returns, while Barra5d aggregates returns over a 5-day horizon - Both models are tested for their ability to generate excess returns relative to the CSI 1000 index [8][31] - **Model Evaluation**: Barra5d demonstrates strong year-to-date performance, significantly outperforming the benchmark, while Barra1d shows consistent but less pronounced gains [8][31] --- Model Backtest Results 1. GRU Model - **Excess Return**: Positive across most stock pools, with occasional underperformance in specific sub-models like `barra5d` [5][6][7] 2. Open1d Model - **Weekly Excess Return**: -0.01% - **Year-to-Date Excess Return**: 5.23% [32] 3. Close1d Model - **Weekly Excess Return**: -0.38% - **Year-to-Date Excess Return**: 3.64% [32] 4. Barra1d Model - **Weekly Excess Return**: 0.65% - **Year-to-Date Excess Return**: 3.80% [32] 5. Barra5d Model - **Weekly Excess Return**: 0.02% - **Year-to-Date Excess Return**: 6.44% [32] --- Quantitative Factors and Construction 1. Factor Name: Beta - **Factor Construction Idea**: Measures historical beta to capture market sensitivity [15] - **Factor Construction Process**: Historical beta is calculated based on the covariance of stock returns with market returns [15] 2. Factor Name: Momentum - **Factor Construction Idea**: Captures historical excess return trends [15] - **Factor Construction Process**: - Momentum = 0.74 * Historical Excess Return Volatility + 0.16 * Cumulative Excess Return Deviation + 0.1 * Historical Residual Return Volatility [15] 3. Factor Name: Volatility - **Factor Construction Idea**: Measures stock price fluctuations to identify high-volatility stocks [15] - **Factor Construction Process**: - Volatility = Weighted combination of historical residual return volatility and other metrics [15] 4. Factor Name: Growth - **Factor Construction Idea**: Focuses on earnings and revenue growth rates [15] - **Factor Construction Process**: - Growth = 0.24 * Earnings Growth Rate + 0.47 * Revenue Growth Rate [15] 5. Factor Name: Liquidity - **Factor Construction Idea**: Measures stock turnover to identify liquid stocks [15] - **Factor Construction Process**: - Liquidity = 0.35 * Monthly Turnover + 0.35 * Quarterly Turnover + 0.3 * Annual Turnover [15] --- Factor Backtest Results 1. Beta Factor - **Weekly Long-Short Return**: Positive [16][18] 2. Momentum Factor - **Weekly Long-Short Return**: Negative [16][18] 3. Volatility Factor - **Weekly Long-Short Return**: Positive [16][18] 4. Growth Factor - **Weekly Long-Short Return**: Positive [16][18] 5. Liquidity Factor - **Weekly Long-Short Return**: Positive [16][18]

Quantitative Models and Construction Methods 1. Model Name: Diffusion Index Model - **Model Construction Idea**: The model is used to monitor the critical points of future diffusion index changes, providing insights into potential market turning points[34][35] - **Model Construction Process**: The diffusion index is calculated based on the relative price changes of constituent stocks over a specific time window. For example, if all constituent stocks drop by 5% after 5 days, the diffusion index value is 0.33. The current diffusion index value is 0.82, indicating a relatively uniform distribution[34][35] - **Model Evaluation**: The model provides a systematic way to observe market heat and potential upward space, though it is sensitive to the dynamic updates of constituent stocks[34][35] 2. Model Name: First Threshold Method (Left-Side Trading) - **Model Construction Idea**: This method triggers a sell signal when the diffusion index reaches a predefined threshold[39] - **Model Construction Process**: The first threshold method triggered a sell signal on May 8, 2025, when the diffusion index closed at 0.9850[39] 3. Model Name: Delayed Threshold Method (Right-Side Trading) - **Model Construction Idea**: Similar to the first threshold method but with a delayed signal to confirm the trend[41][43] - **Model Construction Process**: The delayed threshold method triggered a sell signal on May 15, 2025, when the diffusion index closed at 0.8975[43] 4. Model Name: Dual Moving Average Method (Adaptive Trading) - **Model Construction Idea**: This method uses two moving averages to adaptively identify trading signals[44] - **Model Construction Process**: The dual moving average method issued a sell signal again on August 4, 2025[44] --- Model Backtesting Results 1. Diffusion Index Model - Current diffusion index value: 0.82[34][35] 2. First Threshold Method - Triggered sell signal at diffusion index value: 0.9850[39] 3. Delayed Threshold Method - Triggered sell signal at diffusion index value: 0.8975[43] 4. Dual Moving Average Method - Triggered sell signal on August 4, 2025[44] --- Quantitative Factors and Construction Methods 1. Factor Name: One-Year Volatility Factor - **Factor Construction Idea**: Measures the stock's price volatility over the past year[3][29] - **Factor Construction Process**: Rank IC for this factor this week is 0.135, with a historical average of -0.032[3][29] 2. Factor Name: Residual Volatility Factor - **Factor Construction Idea**: Captures the residual volatility of stock returns after accounting for market movements[3][29] - **Factor Construction Process**: Rank IC for this factor this week is 0.057, with a historical average of -0.039[3][29] 3. Factor Name: Growth Factor - **Factor Construction Idea**: Reflects the growth potential of stocks based on financial metrics[3][29] - **Factor Construction Process**: Rank IC for this factor this week is 0.053, with a historical average of -0.004[3][29] 4. Factor Name: Leverage Factor - **Factor Construction Idea**: Measures the financial leverage of companies[3][29] - **Factor Construction Process**: Rank IC for this factor this week is 0.042, with a historical average of -0.006[3][29] 5. Factor Name: Illiquidity Factor - **Factor Construction Idea**: Captures the illiquidity of stocks based on trading volume and price impact[3][29] - **Factor Construction Process**: Rank IC for this factor this week is 0.041, with a historical average of 0.04[3][29] 6. Factor Name: 10-Day Return Factor - **Factor Construction Idea**: Measures the stock's return over the past 10 days[3][29] - **Factor Construction Process**: Rank IC for this factor this week is -0.131, with a historical average of -0.061[3][29] 7. Factor Name: Nonlinear Market Cap Factor - **Factor Construction Idea**: Captures the nonlinear relationship between market capitalization and stock returns[3][29] - **Factor Construction Process**: Rank IC for this factor this week is -0.13, with a historical average of -0.033[3][29] 8. Factor Name: Logarithmic Market Cap Factor - **Factor Construction Idea**: Uses the logarithm of market capitalization to explain stock returns[3][29] - **Factor Construction Process**: Rank IC for this factor this week is -0.13, with a historical average of -0.033[3][29] 9. Factor Name: 10-Day Total Market Cap Turnover Factor - **Factor Construction Idea**: Measures the turnover of total market capitalization over the past 10 days[3][29] - **Factor Construction Process**: Rank IC for this factor this week is -0.13, with a historical average of -0.06[3][29] 10. Factor Name: PE_TTM Reciprocal Factor - **Factor Construction Idea**: Uses the reciprocal of the price-to-earnings ratio (trailing twelve months) as a valuation metric[3][29] - **Factor Construction Process**: Rank IC for this factor this week is -0.129, with a historical average of 0.017[3][29] --- Factor Backtesting Results Top 5 Factors by Rank IC This Week 1. One-Year Volatility Factor: 0.135[3][29] 2. Residual Volatility Factor: 0.057[3][29] 3. Growth Factor: 0.053[3][29] 4. Leverage Factor: 0.042[3][29] 5. Illiquidity Factor: 0.041[3][29] Bottom 5 Factors by Rank IC This Week 1. 10-Day Return Factor: -0.131[3][29] 2. Nonlinear Market Cap Factor: -0.13[3][29] 3. Logarithmic Market Cap Factor: -0.13[3][29] 4. 10-Day Total Market Cap Turnover Factor: -0.13[3][29] 5. PE_TTM Reciprocal Factor: -0.129[3][29]

Quantitative Models and Construction 1. Model Name: Diffusion Index Industry Rotation Model - **Model Construction Idea**: This model is based on the principle of price momentum, aiming to capture upward trends in industries through a diffusion index[24][25]. - **Model Construction Process**: The diffusion index is calculated for each industry, reflecting the proportion of stocks within the industry that exhibit upward momentum. The index ranges from 0 to 1, where higher values indicate stronger upward trends. The model selects industries with the highest diffusion index values for rotation. - Formula: Not explicitly provided in the report - **Model Evaluation**: The model has shown mixed performance over the years. It performed well in capturing trends during certain periods (e.g., pre-September 2021) but struggled during market reversals or when trends shifted to mean-reversion patterns[24]. 2. Model Name: GRU Factor Industry Rotation Model - **Model Construction Idea**: This model leverages GRU (Gated Recurrent Unit) deep learning networks to process high-frequency volume and price data, aiming to identify industry rotation opportunities[37]. - **Model Construction Process**: The GRU network is trained on historical minute-level data to predict industry factor rankings. The model then selects industries with the highest predicted factor scores for rotation. - Formula: Not explicitly provided in the report - **Model Evaluation**: The GRU model has demonstrated strong adaptability in short-term scenarios but has underperformed in long-term or extreme market conditions. Its reliance on high-frequency data makes it sensitive to market noise[37]. --- Model Backtesting Results 1. Diffusion Index Industry Rotation Model - **Annualized Excess Returns**: - 2021: +25% (pre-September), followed by significant drawdowns later in the year - 2022: +6.12% - 2023: -4.58% - 2024: -5.82% - 2025 (YTD as of August): +2.71%[24][28] - **Monthly Performance (August 2025)**: - Average Return: +4.18% - Excess Return (vs. Equal-Weighted Industry Index): +0.78%[28] 2. GRU Factor Industry Rotation Model - **Annualized Excess Returns**: - 2025 (YTD as of August): -8.59%[31][34] - **Monthly Performance (August 2025)**: - Average Return: +1.80% - Excess Return (vs. Equal-Weighted Industry Index): -1.58%[34] --- Quantitative Factors and Construction 1. Factor Name: Diffusion Index - **Factor Construction Idea**: Measures the proportion of stocks within an industry exhibiting upward momentum, serving as a proxy for industry-level price trends[25]. - **Factor Construction Process**: - The diffusion index is calculated weekly for each industry. - Industries are ranked based on their diffusion index values, with higher values indicating stronger momentum. - Example Rankings (as of August 22, 2025): - Top Industries: Comprehensive Finance (1.0), Comprehensive (1.0), Steel (1.0) - Bottom Industries: Coal (0.262), Electric Utilities (0.587), Real Estate (0.694)[25][26]. 2. Factor Name: GRU Industry Factor - **Factor Construction Idea**: Derived from GRU deep learning models, this factor captures industry-level signals based on high-frequency trading data[37]. - **Factor Construction Process**: - The GRU model processes minute-level volume and price data to generate factor scores for each industry. - Industries are ranked based on their GRU factor scores. - Example Rankings (as of August 22, 2025): - Top Industries: Building Materials (3.32), Electronics (2.36), Non-Banking Finance (1.97) - Bottom Industries: Electric Utilities (-25.33), Banking (-24.29), Pharmaceuticals (-20.97)[32]. --- Factor Backtesting Results 1. Diffusion Index - **Weekly Rankings (August 22, 2025)**: - Top Industries: Comprehensive Finance (1.0), Comprehensive (1.0), Steel (1.0) - Bottom Industries: Coal (0.262), Electric Utilities (0.587), Real Estate (0.694)[25][26]. 2. GRU Industry Factor - **Weekly Rankings (August 22, 2025)**: - Top Industries: Building Materials (3.32), Electronics (2.36), Non-Banking Finance (1.97) - Bottom Industries: Electric Utilities (-25.33), Banking (-24.29), Pharmaceuticals (-20.97)[32].

- The Genie Envisioner platform introduces a video-centric world modeling paradigm, directly modeling robot-environment interactions in the visual space, which retains spatial structure and temporal evolution information. This approach enhances cross-domain generalization and long-sequence task execution capabilities, achieving a 76% success rate in long-step tasks like folding cardboard boxes, outperforming the π0 model's 48%[12][13][16] - The Genie Envisioner platform comprises three core components: GE-Base, a multi-view video world foundation model trained on 3000 hours of real robot data; GE-Act, a lightweight 160M parameter action decoder enabling real-time control; and GE-Sim, a hierarchical action-conditioned simulator for closed-loop strategy evaluation and large-scale data generation[16][17][19] - The GLM-4.5V visual reasoning model, with 106B total parameters and 120B activation parameters, achieves state-of-the-art (SOTA) performance across 41 multimodal benchmarks, including image, video, document understanding, and GUI agent tasks. It incorporates 3D-RoPE and bicubic interpolation mechanisms to enhance 3D spatial relationship perception and high-resolution adaptability[20][21][22] - GLM-4.5V employs a three-stage training strategy: pretraining on large-scale multimodal corpora, supervised fine-tuning with "chain of thought" samples, and reinforcement learning with RLVR and RLHF techniques. This layered training enables superior document processing capabilities and emergent abilities like generating structured HTML/CSS/JavaScript code from screenshots or videos[23][24][26] - VeOmni, a fully modular multimodal training framework, decouples model definition from distributed parallel logic, enabling flexible parallel strategies like FSDP, HSDP+SP, and EP. It achieves 43.98% MFU for 64K sequence training and supports up to 192K sequence lengths, reducing engineering complexity and improving efficiency by over 90%[27][28][31] - VeOmni introduces asynchronous sequence parallelism (Async-Ulysses) and COMET technology for MoE models, achieving linear scalability in training throughput for 30B parameter models under 160K sequence lengths. It also integrates dynamic batch processing and FlashAttention to minimize memory waste and optimize operator-level recomputation[31][32][34] - Skywork UniPic 2.0, a unified multimodal framework, integrates image understanding, text-to-image (T2I) generation, and image-to-image (I2I) editing within a single model. It employs a progressive dual-task reinforcement strategy (Flow-GRPO) to optimize image editing and T2I tasks sequentially, achieving superior performance in benchmarks like GenEval and GEdit-EN[35][38][39] - UniPic 2.0 leverages Skywork-EditReward, an image-editing-specific reward model, to provide pixel-level quality scores. This design enables precise recognition of image elements and generation of corresponding textual descriptions, achieving 83.5 points in MMBench, comparable to 19B parameter models[38][42][43] - FlowReasoner, a query-level meta-agent framework, dynamically generates personalized multi-agent systems for individual queries. It employs GRPO reinforcement learning with multi-objective reward mechanisms, achieving 92.15% accuracy on the MBPP dataset and outperforming baseline models like Aflow and LLM-Blender[63][64][68] - FlowReasoner utilizes a three-stage training process: supervised fine-tuning with synthetic data, SFT fine-tuning for workflow generation, and RL with external feedback for capability enhancement. It demonstrates robust generalization, maintaining high accuracy even when the base worker model is replaced[66][68][69]