Latest news with #SkyworkAI
Yahoo
05-07-2025
- Business
- Yahoo
Skywork-Reward-V2: Leading the New Milestone for Open-Source Reward Models
SINGAPORE, July 5, 2025 /PRNewswire/ -- In September 2024, Skywork first open-sourced the Skywork-Reward series models and related datasets. Over the past nine months, these models and data have been widely adopted by the open-source community for research and practice, with over 750,000 cumulative downloads on the HuggingFace platform, helping multiple frontier models achieve excellent results in authoritative evaluations such as RewardBench. On July 4, 2025, Skywork continues to open-source the second-generation reward models - the Skywork-Reward-V2 series, comprising 8 reward models based on different base models of varying sizes, with parameters ranging from 600 million to 8 billion. These models have achieved top rankings across seven major mainstream reward model evaluation benchmarks. Skywork-Reward-V2 Download Links HuggingFace: GitHub: Technical Report: Reward models play a crucial role in the Reinforcement Learning from Human Feedback (RLHF) process. In developing this new generation of reward models, we constructed a hybrid dataset called Skywork-SynPref-40M, containing a total of 40 million preference pairs. To achieve large-scale, efficient data screening and filtering, Skywork specially designed a two-stage human-machine collaborative process that combines high-quality human annotation with the scalable processing capabilities of models. In this process, humans provide rigorously verified high-quality annotations, while Large Language Models (LLMs) automatically organize and expand based on human guidance. Based on the above high-quality hybrid preference data, we developed the Skywork-Reward-V2 series, which demonstrates broad applicability and excellent performance across multiple capability dimensions, including general alignment with human preferences, objective correctness, safety, resistance to style bias, and best-of-N scaling capability. Experimental validation shows that this series of models achieved the best performance on seven mainstream reward model evaluation benchmarks. 01 Skywork-SynPref-40M: Human-Machine Collaboration for Million-Scale Human Preference Data Screening Even the most advanced current open-source reward models still perform inadequately on most mainstream evaluation benchmarks. They fail to effectively capture the subtle and complex characteristics of human preferences, particularly when facing multi-dimensional, multi-level feedback. Additionally, many reward models tend to excel on specific benchmark tasks but struggle to transfer to new tasks or scenarios, exhibiting obvious "overfitting" phenomena. Although existing research has attempted to improve performance through optimizing objective functions, improving model architectures, and recently emerging Generative Reward Models, the overall effectiveness remains quite limited. We believe that the current fragility of reward models mainly stems from the limitations of existing preference datasets, which often have limited coverage, mechanical label generation methods, or lack rigorous quality control. Therefore, in developing the new generation of reward models, we not only continued the first generation's experience in data optimization but also introduced more diverse and larger-scale real human preference data, striving to improve data scale while maintaining data quality. Consequently, Skywork proposes Skywork-SynPref-40M - the largest preference hybrid dataset to date, containing a total of 40 million preference sample pairs. Its core innovation lies in a "human-machine collaboration, two-stage iteration" data selection pipeline. Stage 1: Human-Guided Small-Scale High-Quality Preference Construction The team first constructed an unverified initial preference pool and used Large Language Models (LLMs) to generate preference-related auxiliary attributes such as task type, objectivity, and controversy. Based on this, human annotators followed a strict verification protocol and used external tools and advanced LLMs to conduct detailed reviews of partial data, ultimately constructing a small-scale but high-quality "gold standard" dataset as the basis for subsequent data generation and model evaluation. Subsequently, we used preference labels from the gold standard data as guidance, combined with LLM large-scale generation of high-quality "silver standard" data, thus achieving data volume expansion. The team also conducted multiple rounds of iterative optimization: in each round, training reward models and identifying model weaknesses based on their performance on gold standard data; then retrieving similar samples and using multi-model consensus mechanisms for automatic annotation to further expand and enhance silver standard data. This human-machine collaborative closed-loop process continues iteratively, effectively improving the reward model's understanding and discrimination of preferences. Stage 2: Fully Automated Large-Scale Preference Data Expansion After obtaining preliminary high-quality models, the second stage turns to automated large-scale data expansion. This stage no longer relies on manual review but uses trained reward models to perform consistency filtering: If a sample's label is inconsistent with the current optimal model's prediction, or if the model's confidence is low, LLMs are called to automatically re-annotate; If the sample label is consistent with the "gold model" (i.e., a model trained only on human data) prediction and receives support from the current model or LLM, it can directly pass screening. Through this mechanism, the team successfully screened 26 million selected data points from the original 40 million samples, achieving a good balance between preference data scale and quality while greatly reducing the human annotation burden. 02 Skywork-Reward-V2: Matching Large Model Performance with Small Model Size Compared to the previous generation Skywork-Reward, Skywork newly released Skywork-Reward-V2 series provides 8 reward models trained based on Qwen3 and LLaMA3 series models, with parameter scales covering from 600 million to 8 billion. On seven mainstream reward model evaluation benchmarks including Reward Bench v1/v2, PPE Preference & Correctness, RMB, RM-Bench, and JudgeBench, the Skywork-Reward-V2 series comprehensively achieved current state-of-the-art (SOTA) levels. Compensating for Model Scale Limitations with Data Quality and Richness Even the smallest model, Skywork-Reward-V2-Qwen3-0.6B, achieves overall performance nearly matching the previous generation's strongest model, Skywork-Reward-Gemma-2-27B-v0.2, on average. The largest scale model, Skywork-Reward-V2-Llama-3.1-8B, achieved comprehensive superiority across all mainstream benchmark tests, becoming the currently best-performing open-source reward model overall. Broad Coverage of Multi-Dimensional Human Preference Capabilities Additionally, Skywork-Reward-V2 achieved leading results in multiple advanced capability evaluations, including Best-of-N (BoN) tasks, bias resistance capability testing (RM-Bench), complex instruction understanding, and truthfulness judgment (RewardBench v2), demonstrating excellent generalization ability and practicality. Highly Scalable Data Screening Process Significantly Improves Reward Model Performance Beyond excellent performance in evaluations, the team also found that in the "human-machine collaboration, two-stage iteration" data construction process, preference data that underwent careful screening and filtering could continuously and effectively improve reward models' overall performance through multiple iterative training rounds, especially showing remarkable performance in the second stage's fully automated data expansion. In contrast, blindly expanding raw data not only fails to improve initial performance but may introduce noise and negative effects. To further validate the critical role of data quality, we conducted experiments on a subset of 16 million data points from an early version. Results showed that training an 8B-scale model using only 1.8% (about 290,000) of the high-quality data already exceeded the performance of current 70B-level SOTA reward models. This result again confirms that the Skywork-SynPref dataset not only leads in scale but also has significant advantages in data quality. 03 Welcoming a New Milestone for Open-Source Reward Models: Helping Build Future AI Infrastructure In this research work on the second-generation reward model Skywork-Reward-V2, the team proposed Skywork-SynPref-40M, a hybrid dataset containing 40 million preference pairs (with 26 million carefully screened pairs), and Skywork-Reward-V2, a series of eight reward models with state-of-the-art performance designed for broad task applicability. We believe this research work and the continued iteration of reward models will help advance the development of open-source reward models and more broadly promote progress in Reinforcement Learning from Human Feedback (RLHF) research. This represents an important step forward for the field and can further accelerate the prosperity of the open-source community. The Skywork-Reward-V2 series models focus on research into scaling preference data. In the future, the team's research scope will gradually expand to other areas that have not been fully explored, such as alternative training techniques and modeling objectives. Meanwhile, considering recent development trends in the field - reward models and reward shaping mechanisms have become core components in today's large-scale language model training pipelines, applicable not only to RLHF based on human preference learning and behavior guidance, but also to RLVR including mathematics, programming, or general reasoning tasks, as well as agent-based learning scenarios. Therefore, we envision that reward models, or more broadly, unified reward systems, are poised to form the core of AI infrastructure in the future. They will no longer merely serve as evaluators of behavior or correctness, but will become the "compass" for intelligent systems navigating complex environments, helping them align with human values and continuously evolve toward more meaningful goals. Additionally, Skywork released the world's first deep research AI workspace agents in May, which you can experience by visiting: Media Contact Company Name: Skywork AI Person: Peter TianEmail: peter@ 2 Science Park DriveCountry: SingaporeWebsite: View original content to download multimedia: SOURCE Skywork AI pte ltd Sign in to access your portfolio
Globe and Mail
23-05-2025
- Business
- Globe and Mail
Skywork.ai Redefines Enterprise Productivity with DeepResearch AI Agents Built for Real-World Industry Demands
In an era where speed alone is no longer enough, emerges as a transformative force—empowering industries with the first AI Office Suite engineered specifically for depth, accuracy, and business impact. Launched globally in May 2025, introduces a new paradigm for professional productivity: AI agents that don't just automate tasks but elevate them—delivering consulting-grade insights, verifiable research, and fully usable outputs tailored to the complex workflows of modern enterprises. From financial services and consulting to education, media, and business, equips organizations to operate faster, think deeper, and execute smarter—turning AI from a tool into a strategic asset. Turning 8 Hours into 8 Minutes The average office worker spends over 60% of their time buried in document creation, data analysis, and presentations—tasks critical to business operations but taxing in both time and cognitive load. new Super Agents are purpose-built to solve this problem. With a single command, users can generate five content formats—documents, slides, spreadsheets, web pages, and even podcasts—reducing hours of manual work to mere minutes, without compromising on accuracy, design, or insight. Purpose-Built Agents That Work the Way You Do empowers users through a suite of five expert agents—each designed to address a specific, high-impact task with professional precision. Rather than offering broad but shallow functionality, delivers deep, task-focused intelligence that aligns with real-world workflows—maximizing output quality, minimizing effort, and accelerating results across every stage of knowledge work. Docs Agent enables users to generate deeply researched, well-structured reports—whether for business strategy, academic publication, or internal planning. It supports advanced reasoning, integrates real-time, traceable sources, and auto-generates visual data charts—helping users craft credible, insight-rich documents in a fraction of the usual time. Slides Agent transforms complex information into compelling, visually engaging presentations. With features like dynamic layouts, integrated videos, and export options to PPTX, PDF, or Google Slides, users can confidently deliver polished decks that communicate clearly and leave a lasting impression. Sheets Agent simplifies data analysis for users of all backgrounds. From descriptive statistics to trend forecasts, it can instantly convert raw datasets into clean, accurate visualizations—such as bar charts, scatter plots, and pie graphs—empowering users to draw insights and make informed decisions with ease. Web Agent gives users the ability to create professional, interactive websites—without writing a single line of code. Whether building a product landing page, internal hub, or event site, users can move from concept to launch in minutes, supported by intuitive structure and responsive design. Podcast Agent turns simple ideas into rich, narrative-driven audio content. Users can generate scripts and fully produced podcast episodes with a single prompt—ideal for educational storytelling, internal communications, or branded content—bringing ideas to life with clarity and emotion. Each agent works seamlessly within unified platform, ensuring outputs are not only fast and flexible but also exportable, editable, and ready for immediate professional use. Together, they empower users to achieve more—across formats, functions, and industries. DeepResearch: Giving Users Unmatched Depth, Clarity, and Confidence At the core of is DeepResearch—a breakthrough engine that redefines what users can expect from AI-generated content. Unlike typical RAG systems that rely on shallow surface-level retrieval, DeepResearch dives up to 10 times deeper into information sources, uncovering richer context, more accurate data, and high-value insights that mirror the quality of expert research. For users, this means generating content that isn't just fast—it's credible, comprehensive, and strategically useful. Whether you're drafting a market analysis, writing an academic paper, or preparing a high-stakes business proposal, DeepResearch ensures your output is backed by validated sources, nuanced synthesis, and professional-grade logic. You don't have to worry about hallucinations or vague generalizations—because every claim can be traced to its origin, and every section is shaped by a real understanding of the topic. With DeepResearch, users gain a powerful advantage: content that stands up to scrutiny, drives decisions, and earns trust. Smarter Inputs, Better Results — Without the Guesswork helps users get exactly what they need—faster and with less trial and error—through its intelligent Clarification Card system. Instead of forcing users to craft perfect prompts or rely on vague instructions, the platform actively guides them to define their goals, context, and constraints upfront. This means you spend less time clarifying, rewriting, or fixing outputs—and more time acting on results that are aligned with your actual needs. Whether you're preparing a report, presentation, or data analysis, ensures the AI understands what you're aiming for before generating anything. The result? Higher accuracy, fewer revisions, and content that fits your workflow from the start. It's not just smarter AI—it's a system that thinks with you, not just for you. All-in-One Multimodal Creation — Powered by Your Ideas gives users the freedom to create far beyond traditional documents and spreadsheets. With its general agent and flexible plug-in ecosystem (MCPs), you can produce engaging videos, voiceovers, music, audiobooks, picture books, and more—all from one streamlined interface. Whether you're developing marketing content, educational media, or creative storytelling, brings multimodal creation into reach—no special tools or technical skills required. Just describe your idea, and the platform helps you bring it to life in the format that fits your goals. Even better, supports a personal knowledge base, allowing you to upload and reuse your own materials—so every piece of content you create is grounded in your existing knowledge, brand voice, or style. This not only saves time, but helps you build smarter, more consistent content over time. Affordable Excellence, Unmatched Impact Unlock the power of enterprise-grade AI—without the enterprise price tag. At just $19.99 in the USA , delivers professional-level performance at a fraction of the cost. With general tasks starting for Cheaper, you get access to deep research, multimodal creation, and expert-level outputs Whether you're a solo creator, agile startup, or global enterprise, scales with you—empowering you to move faster, work smarter, and create with confidence. isn't just another AI tool—it's your productivity partner. Ready to upgrade your workflow, enhance your content, and lead with intelligence? Start your journey today at: Follow us for the latest innovations: X (Twitter): @Skywork_ai YouTube: SkyworkAI GitHub: SkyworkAI Experience the future of work—smarter, faster, and more powerful than ever. Media Contact Company Name: Skywork Contact Person: Jane Wang Email: Send Email Country: Singapore Website:
Yahoo
19-02-2025
- Entertainment
- Yahoo
SkyReels Open Sources the World's First Human-Centric Video Foundation Model for AI Short Drama Creation – SkyReels-V1, Reshaping the AI Short Drama Landscape
Singapore, Feb. 19, 2025 (GLOBE NEWSWIRE) -- On February 18, SkyReels open-sourced the world's first human-centric video foundation model for AI short drama creation, SkyReels-V1, and the world's first SOTA-level expressive portrait image animation based on video diffusion transformers, SkyReels-A1. Open-Source Repositories: SkyReels-V1https:// SkyReels-A1https:// Technical Report: Official Website: Addressing global pain points in AI video generation—such as closed-source models, limited accessibility, high costs, and usability issues—SkyReels is breaking new ground by open-sourcing two SOTA-level models and algorithms, SkyReels-V1 and SkyReels-A1. These cutting-edge technologies for AI short drama creation are now offered to the open-source community and AIGC users. The production format for AI videos and short dramas has been market-validated. SkyReels helps address the challenges in traditional short drama production—such as complex offline processes including scriptwriting, casting, set design, storyboard creation, filming, and post-production, which require substantial manpower, incur high costs, and extend production cycles. 01SkyReels-V1: Human-Centric Video Foundation Model, the world's first open-source video generation model dedicated to AI short drama creation AI short dramas require precise control over both cognitive and physical expressions—integrating lip-sync, facial expression, and body movement generation into a unified process. Currently, lip-sync generation is particularly well-developed, owing to its strong mapping with audio cues that enable high precision and superior user experience. Yet, the true quality of AI short drama generation lies in the nuances of character performance. To dramatically enhance the controllability of facial expressions and body movements, SkyReels-V1 not only meticulously annotates performance details but also processes emotions, scene context, and acting intent, fine-tuning on tens of millions of high-quality, Hollywood-level data points. Research team has implemented advanced technical upgrades to capture micro-expressions, performance subtleties, scene descriptions, lighting, and composition. As a result, characters generated by SkyReels now exhibit remarkably precise acting details—approaching an award-winning level. SkyReels-V1 delivers cinematic-grade micro-expression performance, supporting 33 nuanced facial expressions and over 400 natural motion combinations that faithfully reproduce genuine human emotional expression. As demonstrated in the accompanying video, SkyReels-V1 can generate expressions ranging from hearty laughter, fierce roars, and astonishment to tears—showcasing rich, dynamic performance details. Moreover, SkyReels-V1 brings cinematic-level lighting and aesthetics to AI video generation. Trained on Hollywood-level high-quality film data, every frame generated by SkyReels exhibits cinematic quality in composition, actor positioning, and camera angles. Whether capturing solo performance details or multi-character scenes, the model now achieves precise expression control and high-quality visuals. Importantly, SkyReels-V1 supports both text-to-video and image-to-video generation. It is the largest open-source video generation model supporting image-to-video tasks at equivalent resolution, achieving SOTA-level performance across multiple metrics. Figure 1: Comparison of Text-to-Video Metrics for SkyReels-V1 (Source: SkyReels) Such SOTA-level performance is made possible not only by SkyReels self-developed high-quality data cleaning and manual annotation pipeline—which has built a tens-of-millions–scale dataset from movies, TV shows, and documentaries—but also by 'Human-Centric' multimodal video understanding model, which significantly enhances the ability to interpret human-related elements in video, particularly through in-house character intelligence analysis system. In summary, thanks to the robust data foundation and advanced character intelligence analysis system, SkyReels-V1 can achieve: Cinematic Expression Recognition: 11 types of facial expression understanding for characters in film and drama, including expressions such as disdain, impatience, helplessness, and disgust, with emotional intensity levels categorized into strong, medium, and weak. Character Spatial Awareness: Leveraging 3D reconstruction technology to comprehend spatial relationships among multiple characters, enabling cinematic positioning. Behavioral Intent Understanding: Constructing over 400 behavioral semantic units for precise action interpretation. Scene-Performance Correlation: Analyzing the interplay between characters, wardrobe, setting, and plot. is not only among the very few open-source video foundation models worldwide, but it is also the most powerful in terms of performance for character-driven video generation. With SkyReels self-developed inference optimization framework 'SkyReels-Infer,' the inference efficiency has been significantly improved —achieving 544p video generation on a single RTX 4090 in just 80 seconds. The framework supports distributed multi-GPU parallelism, Context Parallel, CFG Parallel, and VAE Parallel. Furthermore, by implementing FP8 quantization and parameter-level offload, it meets the requirements of low-memory GPUs; support for flash attention, SageAttention, and model compilation optimizations further reduces latency; and leveraging the open-source diffuser library enhances 2: Using equivalent RTX 4090 resources (4 GPUs), the SkyReels-Infer version reduces end-to-end latency by 58.3% compared to the HunyuanVideo official version (293.3s vs. 464.3s). Figure 3: Under similar A800 resource conditions, the SkyReels-Infer version reduces end-to-end latency by 14.7%–28.2% compared to the HunyuanVideo official version, demonstrating a more robust multi-GPU deployment strategy. 02SkyReels-A1: The First SOTA-Level Expressive Portrait Image Animation Algorithm Based on Video Diffusion Transformers To achieve even more precise and controllable character video generation, the SkyReels is open-sourcing SkyReels-A1, a SOTA-level algorithm based on video diffusion transformer for expression and action control. Comparable to Runway Act-One, SkyReels-A1 supports video-driven, film-grade expression capture, enabling high-fidelity micro-expression reproduction. SkyReels-A1 can generate highly realistic and consistency videos for characters in any reference conditions—from portrait of half-body to full-body shots. It achieves a precise simulation of facial expressions, emotional nuances, skin textures, and body movements. By inputting both a reference and a driving video, SkyReels-A1 'transplants' the facial expressions and actions details from the driving video onto the character in the reference image. The resulting video shows no distortion and faithfully reproduces the micro-expressions and body movements from the driving video, even surpassing the video quality generated by Runway Act-One in evaluation. More encouragingly, SkyReels-A1 not only supports profile-based expression control but also enables highly realistic eyebrow and eye micro-expression alignment, along with more pronounced head movements and natural body motions. For example, in the same dialogue scene, while the character generated by Runway Act-One shows noticeable distortion and deviates from the original appearance, SkyReels-A1 preserves the character's details, maintaining authentic nuance and seamlessly blending facial expressions as well as body movements. Furthermore, SkyReels-A1 can drive more dramatic facial expressions scenes. Compared to Runway Act-One that fails to generate the desired effect, SkyReels-A1 support to transfer more complex expression dynamics, enabling the character's facial emotions to naturally synchronize with body movements and scene content for an exceptionally life–like performance. 03Empowering the Global AI Short Drama Ecosystem through Open-Sourcing Video generation models are among the most challenging components of AI short drama creation. Although model generation capabilities have significantly improved over the past year, there remains a considerable gap—particularly given the high production costs. By open-sourcing our SOTA-level models, SkyReels-V1 and SkyReels-A1, SkyReels becomes the first in the AI short drama industry to take such a step. This initiative not only represents a modest yet significant contribution to the industry but also marks a major leap toward fostering a flourishing ecosystem for AI short drama creation and video generation. It's believed that with further advancements in inference optimization and the open-sourcing of controllable algorithms, these models will soon provide users with more cost-effective and highly controllable AIGC capabilities. SkyReels aims to empower users to create AI short dramas at minimal cost, overcome current issues of inconsistent video generation, and enable everyone to generate detailed, controllable character performances using their own computers. This open-sourcing of our video generation models is not only a technological breakthrough that helps narrow the digital divide in the global content industry, but it is also a revolution in cultural production capacity. In the future, the convergence of short dramas, gaming, virtual reality, and other fields will accelerate industrial integration. AI short dramas have the potential to evolve from a 'tech experiment' into a mainstream creative medium and become a new vehicle for global cultural expression. 'Achieve artificial general intelligence and empower everyone to better shape and express themselves.' With this open-source initiative, SkyReels will continue to release more video generation models, algorithms, and universal models—advancing AGI equity and fostering the sustained growth and prosperity of the AI short drama ecosystem, while benefiting the open-source community, developer ecosystems, and the broader AI industry. CONTACT: Jingnan Fu Skywork AI fujingnan(at) in to access your portfolio
