qwen

The Slow Fog security team recently disclosed the discovery of a malicious GitHub repository impersonating the "Qwen 3.8 27B local quantitative model." The repository claims that the model file should exceed 16GB, while the actual downloaded content is only about 487KB, containing disguised files, a LuaJIT interpreter, and obfuscated Lua scripts. Slow Fog emphasizes that the official Qwen project has not been compromised.Slow Fog's analysis indicates that after the malicious program runs, it will collect host information, capture the screen, and send it to the attacker's C2 server; when the hard-coded server fails, it will also read backup C2 addresses from contracts on the Polygon chain, allowing the attacker to rotate infrastructure through on-chain transactions.Subsequent payloads can steal browser login information, cookies, history, email, WinSCP, Steam credentials, as well as files and extension data related to cryptocurrency wallets. Slow Fog also found that at least 23 GitHub repositories and 29 similar compressed packages used the same Lua delivery chain.

On August 27, B.AI announced that the latest open-source Qwen3.8-Flash model from Alibaba Tongyi Qianwen has officially landed on the platform, with the API official group fully open for free calls. This multimodal MoE model serves as a forward-looking preview of the Qwen4 architecture, featuring a total of 125B parameters and 51B N-gram Embedding, activating 6B parameters per token.The model integrates the GDN and QSA mixed attention mechanisms for the first time, natively supporting 262K context and can be extended to 1M via YaRN. Developers can start calling it at zero cost for tasks such as programming assistance, agent workflows, and document analysis.

According to Jinshi, Alibaba Cloud announced that the billing unit price for the Qwen3.8-Flash model mode on the Bailian large model service platform will be adjusted downwards starting from 12:00:00 Beijing time on August 27, 2026. Qwen3.8-Flash is the latest multimodal large model launched by Qianwen, natively supporting a million-level context window, capable of processing ultra-long documents, code repositories, and complex dialogues in one go. It performs outstandingly in scenarios such as programming assistance, intelligent agent collaboration, and image-text understanding, and is compatible with OpenAI and Anthropic interface protocols, making it suitable for building high-concurrency applications and intelligent workflows.The specifics are as follows: the input price before adjustment was 1 yuan, and after adjustment it will be 0.8 yuan; the output price before adjustment was 3 yuan, and after adjustment it will be 2.7 yuan.

Ali Qianwen released the Qwen3.8-Flash model, which is described as a multimodal MoE model and an early preview of the Qwen4 architecture.The production version of Qwen3.8-Flash will soon be available through the Qwen Cloud API, with a price of only $0.16 per million input tokens and $0.47 per million output tokens. The model has 125 billion parameters + 51 billion N-gram embedding parameters, but only activates 6 billion parameters per token, achieving extremely high cost-effectiveness.

The Alibaba Qwen team will release Qwen 3.8-Flash-Next on Wednesday, which is a mixture of experts (MoE) model with a total of 125 billion parameters, but only 6 billion parameters are activated for each token. The team positions it as a preview version of the next-generation Qwen 4 architecture, rather than the final flagship model.The model is described as multimodal and is built on the upcoming Qwen 4 architecture. The team stated that the early release of this version is to prepare developers for the subsequent complete Qwen series. The model weights will be hosted on the Hugging Face and ModelScope platforms.As of now, Qwen has not disclosed the benchmark results for this model, nor has it released comparative data with its own Qwen 3 series or overseas competitors. The specific figures of 125 billion and 6 billion parameters have not been officially confirmed. As an open-weight model, developers can download, fine-tune, and run it for free without sending data to a closed API, which helps reduce the cost of hosting the model.

According to Business Insider, Thomson Reuters launched its first self-developed AI model, Thomson-1, on Monday. This model is based on Snowdon and constructed through "re-alignment" of Alibaba's open-source Qwen model. Open-source means anyone can download and modify the model for free. Thomson-1 will take over some tasks previously handled by Anthropic's Claude, but it is not intended to completely replace collaboration with Anthropic and other labs, initially focusing on the company's areas of expertise, starting with document review.This move aims to address the high AI costs brought by models like Claude and OpenAI Codex, and reflects the use of cheaper Chinese open-source AI. CTO Joel Hron stated that the main reason is to better leverage Thomson Reuters' own expertise and control costs. The company expanded its collaboration with Anthropic in May this year for the AI legal assistant CoCounsel, which still primarily relies on Claude. Hron said, "Our main goal is to gradually make Thomson the model that drives more and more capabilities for CoCounsel."Thomson Reuters, in collaboration with a team from Imperial College London, spent months transforming Qwen to build Snowdon and ensured it is "ethically and politically bias-processed and safe to use." Hron pointed out that having a proprietary model allows for development based on its own intellectual property rather than continuously paying external AI companies, comparing it to renting versus buying a house: renting provides shelter but does not accumulate long-term equity.

According to "DuJia," DingTalk is advancing a brand new AI office application QwenNote (Qianwen Listening Note). This application is positioned as an AI personal assistant, integrating real-time voice transcription, summarization, and translation through a combination of software and hardware, and deeply integrating with AI Agent, embedding Agent capabilities into voice input, promoting a shift from simple recording to automated execution. The application supports real-time transcription and bilingual recognition in Chinese and English, as well as language switching, and can generate structured meeting minutes, outlines, and to-do lists, with built-in AI Q&A and quick commands based on listening materials.QwenNote offers a voice memo function, requiring a QR code scan to connect to the recording device. By long-pressing the button on the back of the device, users can record inspirations, and once the recording is complete, it will automatically archive, generate a title and brief summary, and timestamp it. The product also features a stealth protection mode, which, when activated, will physically delete the original audio and only retain the transcribed text to accommodate confidential scenarios. The associated hardware QwenNote A2 has already been showcased, which is part of the Qianwen Listening Note hardware ecosystem, which also includes DingTalk A1, DingTalk A1 Pro, Cleer H1, etc., and users can complete the binding by scanning a QR code.Reports indicate that DingTalk hopes to complement the offline voice collection entry through the integration of software and hardware, forming a closed loop of on-site audio collection, real-time bilingual transcription, AI meeting minutes Q&A, and DingTalk organizational collaboration. Listening materials can be synchronized to DingTalk AI Listening Note and support personal private isolation. On the software side, it continues to embed large models into documents, meetings, and IM scenarios, while on the hardware side, it expands the Qianwen Listening Note product line. Relevant hardware has not yet been widely publicly searched for more formal release information.

B.AI welcomes several major updates this week. In terms of benefits, new users can log in with Bitget Wallet, Binance Wallet, or imToken Wallet to instantly receive 1 million free Credits; logging in with an invitation code grants an additional 300,000, with a maximum cumulative total of 1.3 million when combined with the exclusive wallet login gift. Referrers also enjoy permanent rebates on friend deposits and subscriptions. Recharge discounts are released simultaneously, with a 1:1 equivalent quota gift exclusive to the BNB Chain channel, while various other payment methods enjoy a 1:0.5 rebate, with a maximum of $100 worth of points available per user.On the model side, there is another breakthrough, as the "Self-selected Service Provider" lineup is significantly expanded this week, adding the Nebula channel and launching historic calling discounts as low as 10%; at the same time, Alibaba's Tongyi Qianwen flagship model Qwen3.8-Max officially lands on the platform, now open for a limited-time free experience. Currently, the self-selected mode has fully covered global mainstream large model series such as Claude, GPT, and Gemini, combined with recharge gifts and multiple discount tiers, continuously releasing extreme computing power cost-effectiveness for global developers and enterprise users.

According to Reuters, Alibaba plans to require its next-generation Qwen open-source AI model's heavy commercial users to share a portion of their revenue with it. This initiative is similar to the approach taken by Moonlight Dark Side with Kimi K3: the licensing terms for Kimi K3 stipulate that if the model is sold as a service and the annual revenue exceeds $20 million, a commercial agreement must be negotiated with Moonlight Dark Side, with reports suggesting a revenue-sharing ratio of up to 30%. The specific revenue-sharing ratio for Alibaba is still under discussion.The report points out that such revenue-sharing agreements have gradually taken shape between Chinese AI companies and American cloud platforms. Several American cloud providers, including DigitalOcean, have signed commercial agreements with Moonlight Dark Side. In terms of pricing, the input/output token price for Kimi K3 is about one-third that of the Anthropic Fable model. Additionally, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has also joined the open-source camp and released its first open-source model last month. This move signifies that Chinese AI companies are exploring sustainable commercialization paths on open-source models through a "free open-source + commercial charging" freemium model.

The latest generation of Alibaba's Tongyi Qianwen 24 trillion parameter flagship model Qwen3.8-Max is officially launched on the B.AI platform. This model has achieved comprehensive improvements in programming, real office work, long-term tasks, and multimodal intelligence agents, and is currently available for a limited-time free experience.To celebrate the launch, B.AI is offering multiple benefits: new users who log in using Binance Wallet, Bitget Wallet, or imToken will receive 1 million free Credits; recharge to enjoy high rewards, with a 1:1 bonus for BNB Chain recharges, and 1:0.5 for other methods, with a maximum additional reward of 100 USD per user; invite friends to register, and friends will receive 300,000 Credits, with a cumulative maximum of 1.3 million when combined with login gifts, while the inviter will also enjoy rewards from friends' recharges and subscriptions.

The reporter learned that the Alibaba Zhenwu M890 super node has successfully adapted to Qwen3.8 and has launched on the Alibaba Cloud Bailian platform to provide model inference services.Alibaba Cloud claims that this is the first super node in the country to successfully run a large model with over 20 trillion parameters.Qwen3.8 is Alibaba's latest flagship model, with a parameter scale of up to 24 trillion. The Zhenwu M890 is the new generation training and inference integrated AI chip from PingTouGe. The Zhenwu super node instance connects 64 Zhenwu M890 chips through the ICN Switch 1.0 interconnection chip, achieving high-speed interconnection of 800GB/s, with a video memory scale reaching 9TB, allowing a single instance to support the expert parallel requirements of a large-scale 20 trillion parameter MoE model.

According to The Information, sources revealed that Tencent has invested in a new AI laboratory founded by Lin Junyang, the former core leader of Alibaba's Tongyi Qwen large model. In Lin Junyang's first round of financing, Tencent invested $20 million, with the total amount raised in the first round reaching several hundred million dollars, and a post-investment valuation of approximately $2 billion. The first round of financing was led by Sequoia China and Gaorong Capital, with both parties investing $100 million each. After completing the first round of transactions, Lin Junyang is already seeking the next round of financing.Before this transaction, Tencent had already invested in several local large model developers in China, including Dark Side of the Moon, MiniMax, and Zhiyu AI, and is currently participating in the latest round of financing for DeepSeek. Lin Junyang left Alibaba in early March this year. Due to disagreements with Alibaba regarding team restructuring, he chose to publicly announce his resignation on the social platform X.

Ali Qianwen announces the open-source Qwen-Scope------an interpretability module trained based on the Qwen3 series and Qwen3.5 series models, with application scenarios including directed control of inference results, data classification and synthesis, model training and optimization, evaluation sample distribution analysis and comparison, etc.The weights of this open-source Qwen-Scope involve 7 large models, covering the dense models and mixture of experts models of the Qwen3 and Qwen3.5 series, with a total of 14 groups of sparse autoencoder weights.

Ethereum co-founder Vitalik Buterin praised Alibaba's Qwen 35B parameter model's problem-solving ability in a social media post. This artificial intelligence model successfully solved a complex mathematical problem, consuming about 15,000 tokens during the reasoning process, and completed the calculation in just 3 minutes on a local device.Vitalik Buterin pointed out that the problem demonstrated the technique of embedding modular arithmetic into continuous function expressions using trigonometric functions, showcasing the Qwen model's capabilities in mathematical reasoning.

"AI Trading Competition"All six models have incurred losses, with QWEN3 MAX currently in the lead with an account value of $9730, a total loss of $269.69, and a trading win rate of 30.2%. This model set records for the largest single profit of $8176 and the largest single loss of $1728, completing a total of 43 trades with trading fees of $1654.The second-ranked DEEPSEEK CHAT V3.1 has an account value of $6562, a total loss of $3438, and a trading win rate of 24.4%. This model achieved a single maximum profit of $7378 and a single maximum loss of $1224, completing a total of 41 trades with trading fees of $690.30.GEMINI 2.5 PRO ranks third, with an account value of $5906, a total loss of $4094, and a trading win rate of 25.6%. This model conducted 238 trades, with trading fees of $1331, a single maximum profit of $496.82, and a single maximum loss of $750.02.The remaining three models performed as follows: GROK 4 with an account value of $4614 and a total loss of $5386; CLAUDE SONNET 4.5 with an account value of $4471 and a total loss of $5529; and GPT 5 with an account value of $4078 and a total loss of $5922.

According to data from Alpha Arena, the first nof1 AI model trading competition has concluded.Among them, Qwen3 Max ranked first, with a return rate of 22.3%, a win rate of 30.2%, a total profit and loss of $2,232, and a total of 43 trades;DeepSeek Chat V3.1 ranked second, with a return rate of 4.89%, a win rate of 24.4%, a total profit and loss of $489.08, and a total of 41 trades.The remaining models all suffered significant losses, with Claude Sonnet 4.5 losing 30.81%, Grok 4 losing 45.3%, Gemini 2.5 Pro losing 56.71%, and GPT 5 losing 62.66%.

ChainCatcher message, as the cryptocurrency market continues to experience significant volatility, the performance of various models in the "AI Trading Competition" has collectively declined. Among the six major AI models, only DeepSeek and QWEN3 are currently in a profitable state, while the accounts of other AI models are all in a loss state. The current ranking by total account amount is as follows:DeepSeek (15,339.29 USDC, profit of 5,339.29 USDC)Qwen3 (12,378.47 USDC, profit of 2,378.47 USDC)Claude (99,329.63 USDC, loss of 670.37 USDC)Grok (6,917.10 USDC, loss of 3,082.90 USDC)Gemini (3,603.72 USDC, loss of 6,396.28 USDC)GPT5 (2,820.99 USDC, loss of 7,179.01 USDC)

ChainCatcher news, the decentralized AI training platform FLock.io officially announces a strategic partnership with Qwen, a leading series of open-source large language models under Alibaba Cloud, marking a deep connection between decentralized AI and blockchain technology within the mainstream AI ecosystem.This collaboration focuses on three major technological breakthroughs:Technological Integration: Combining Alibaba Cloud's centralized infrastructure with FLock.io's decentralized technology to jointly develop domain-specific and general AI models, while promoting the seamless integration of decentralized AI models into centralized platforms.Data Privacy Protection: Exploring the combination of distributed ledger technology and federated learning to address data privacy and sovereignty issues in model training, providing innovative solutions for the secure application of private data.Collaborative Innovation: Through joint research and technological collaboration, creating a more inclusive, scalable, and privacy-preserving AI ecosystem, facilitating the collaborative development of centralized and decentralized AI.As one of the world's leading large language models, Qwen has demonstrated outstanding performance in multiple authoritative benchmark tests and is widely used in natural language processing, content generation, and other fields. Through this collaboration, FLock and Qwen will work together to explore deeper technological integration, maintaining the advantages of high-performance AI models while promoting the practical application of decentralized AI training in a broader range of scenarios, making it more accessible, flexible, and valuable in both centralized and decentralized ecosystems.