Alibaba's Qwen3 Challenges AI Giants with Hybrid Reasoning Breakthrough
The Qwen3 series comprises six dense models and two Mixture-of-Experts models, ranging from 0.6 billion to 235 billion parameters. Notably, the Qwen3-235B-A22B model, with 22 billion active parameters, has achieved leading scores in various evaluations. On the Codeforces platform, it attained an Elo rating of 2056, outperforming DeepSeek-R1 and Google's Gemini 2.5 Pro. In the AIME'24 and AIME'25 math benchmarks, it scored 85.7 and 81.4, respectively, indicating strong mathematical reasoning capabilities.
A distinctive feature of Qwen3 is its hybrid reasoning capability, allowing dynamic switching between 'thinking' and 'non-thinking' modes. The 'thinking' mode is designed for complex, multi-step tasks such as mathematics, coding, and logical deduction, while the 'non-thinking' mode facilitates fast, general-purpose responses. This dual-mode operation aims to optimize performance across a broad range of applications.
The models have been trained on 36 trillion tokens across 119 languages and dialects, enhancing their multilingual capabilities. This extensive training dataset contributes to Qwen3's proficiency in handling diverse linguistic contexts, making it a versatile tool for global applications.
In terms of deployment efficiency, Qwen3-235B-A22B offers significant advantages. Despite its large parameter count, the model activates only 22 billion parameters during inference, reducing computational requirements and associated costs. This efficiency is particularly beneficial for developers seeking high-performance models without prohibitive resource demands.
The Qwen3 series is available under the Apache 2.0 license, promoting accessibility and collaboration within the AI research community. Developers can access the models through platforms such as Hugging Face and GitHub, facilitating integration into various applications, including mobile devices, smart glasses, autonomous vehicles, and robotics.
Alibaba's release of Qwen3 represents a strategic move to position itself as a leader in the AI domain, challenging established players like OpenAI and Google. The company's focus on hybrid reasoning and multilingual support reflects a commitment to advancing AI capabilities in a manner that is both innovative and inclusive.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Channel Post MEA
a day ago
- Channel Post MEA
Oracle Integrates Google's Gemini AI
Oracle and Google Cloud have expanded their partnership to offer customers access to Google's most advanced AI models, starting with Gemini 2.5, via Oracle Cloud Infrastructure (OCI) Generative AI service. Oracle customers can now utilize the latest Gemini models to build AI agents for a wide range of use cases including multimodal understanding, advanced coding and software development tasks, productivity and workflow automation, and research and knowledge retrieval. Oracle plans to make Google's entire range of Gemini models available via OCI Generative AI service through new integrations with Vertex AI, including cutting edge models for video, image, speech, and music generation and specialized industry models like MedLM. In the future, Oracle will collaborate with Google Cloud to make Gemini models via Vertex AI available as an option within Oracle Fusion Cloud Applications, providing customers with a broader choice to enhance workflows in finance, HR, supply chain, sales, service, and marketing. Oracle customers can use their existing Oracle Universal Credits to start leveraging Google's Gemini models. 'Today, leading enterprises are using Gemini to power AI agents across a range of use cases and industries,' said Thomas Kurian, CEO, Google Cloud. 'Now, Oracle customers can access our leading models from within their Oracle environments, making it even easier for them to begin deploying powerful AI agents that can support developers, streamline data integration tasks, and much more.' Google's Gemini models excel in enterprise use cases thanks to their ability to ground responses in up-to-date Google Search data for accuracy, large context windows, strong encryption and data privacy policies, and leading reasoning abilities. 'Oracle has been intentional in offering model choice curated for the enterprise, spanning open and proprietary models,' said Clay Magouyrk, president, Oracle Cloud Infrastructure. 'The availability of Gemini on OCI Generative AI service highlights our focus on delivering powerful, secure, and cost-effective AI solutions that help customers drive innovation and achieve their business goals.' Oracle brings leading-edge AI technology close to enterprise data and prioritizes security, adaptability, and scalability. This helps customers across industries apply the right AI technologies, including generative and agentic AI, to the right business scenarios for immediate results. In addition, thousands of AI innovators are leveraging OCI's cost-effective, purpose-built AI capabilities to run the most demanding AI workloads faster. OCI bare metal GPU instances can power applications for generative AI, natural language processing, computer vision, and recommendation systems.
Channel Post MEA
a day ago
- Channel Post MEA
OPSWAT Report Reveals 127% Rise In Malware
OPSWAT has released its first-ever Threat Landscape Report , revealing key insights from over 890,000 sandbox scans in the last 12 months. This report offers a unique lens into the evolving nature of cyberthreats. The findings are clear: traditional detection methods are being outpaced, with a 127% rise in malware complexity and a staggering 1 in 14 files—initially deemed 'safe' by legacy systems—proven to be malicious. This report is a call to action for industries relying on outdated defenses and the importance of multi-layered solutions. Key Findings: 127% Increase in Malware Complexity Behavioral telemetry revealed a 127% rise in multi-stage malware complexity over the past year. OPSWAT's sandbox uncovered layered threats designed to evade analysis, including obfuscated loaders such as NetReactor and evasive behaviors missed by traditional tools. These results show that modern malware intends to confuse, not flood, which is why OPSWAT's pipeline is purpose-built to unpack that complexity. Proactive Threat Detection OPSWAT analysis reclassified 7.3% of files that were silent across open-source intelligence (OSINT) feeds as malicious, on average 24 hours earlier than public data sources. These were confirmed executions, not speculative flags, highlighting how adaptive analysis can close dangerous gaps left by static and reputation-based systems. Campaign-Level Threat Correlation With 890,000+ sandbox scans, OPSWAT connects the dots across threats. It identifies shared TTPs, reused C2 infrastructure, and behavioral patterns across campaigns. This provides defenders with context-rich, actionable intelligence instead of noisy indicators. 99.97% Detection Accuracy OPSWAT's behavioral and machine learning pipeline delivers results. Aided by a newly enhanced PE emulator, the platform identified sophisticated threats such as: Clipboard hijacking via ClickFix Steganography-wrapped loaders C2 channels embedded in Google services .NET Bitmap malware loaders delivering Snake Keylogger payloads 'Our strength lies in precision, behavioral depth, and early visibility into emerging attacks,' said Jan Miller, Chief Technology Officer of Threat Analysis at OPSWAT. 'That's what sets OPSWAT apart in delivering high-fidelity, context-aware threat intelligence.' Why It Matters As critical infrastructure, government systems, and enterprise networks face growing targeting from increasingly modular and evasive malware, the findings of this report spotlight the evolving adversary playbook and the need for integrated, multilayered solutions. Cybersecurity leaders must now prioritize adaptability, shared intelligence, reassessing technology, and fast behavioral detection pipelines to protect systems from known threats, but also to keep pace with a rapidly evolving threat landscape and whatever is on the horizon. part of the OPSWAT MetaDefender Platform, powers advanced threat detection and file analysis across critical environments. Download the full report and learn more about OPSWAT's integrated pipeline at .
Arabian Business
a day ago
- Arabian Business
Saudi's PIF sold its entire holding in Meta, Shopify, FedEx and Alibaba in Q2
Saudi Arabia's Public Investment Fund (PIF) has sold all the shares it held in Meta, Shopify and PayPal in the second quarter of the year, according to its latest 13F filing with the US Securities and Exchange Commission (SEC). PIF also sold its stakes in Chinese online retail giant Alibaba Group, Brazil's banking company Nu Holdings, and US logistics company FedEx. According to Reuters, PIF's total exposure to US equities, which include call options that give the state investor a right to buy an underlying asset at a specified price within a specific time period, was valued at US$23.8 billion at the end of Q2, versus US$25.5 billion at the end of the first quarter – a decline of US$1.7 billion. At the end of March, PIF held 667,996 class A shares in Meta, 1.25 million class A shares in Shopify, 1.76 million shares in PayPal, 6.83 million class A shares in Nu Holdings, 1.61 million ADS in Alibaba, and 498,164 common shares in FedEx. In terms of value, PIF's most significant holdings are in two mobility and a gaming company – ride-hailing Uber, EV maker Lucid and Electronic Arts. It held 72.84 million shares in Uber, valued at US$6.8 billion, and 1.77 billion shares in Lucid for a total value of US$3.74 billion. In EA, PIF held 24.8 million shares for a total value of US$3.96 million.
