Nvidia's NAS-Tuned Model Triples Throughput, Bragi Puts GPT-4.1 in Earbuds, and Alibaba Sunsets Virtual Numbers
Hello, developer friends!
Here is the RTE Developer Daily, bringing you news and gossip every day. Our community editorial team curates and shares 'topical technology', 'highlighted products', 'thought-provoking articles', 'opinionated viewpoints', and 'noteworthy events' in the RTE (Real-Time Engagement) field. However, the content only represents the editor's personal views. Everyone is welcome to leave comments, follow threads, and join discussions.
Editors for this issue: @Zhao Yiling, @Bob
01 Topical Technology
1. Nvidia's New Open-Source Model: Triple Throughput, Runs on a Single Card, and Achieves Inference SOTA
Llama Nemotron Super v1.5 is short for Llama-3.3-Nemotron-Super-49B-V1.5. It is an upgraded version of Llama-3.3-Nemotron-Super-49B-V1 (a derivative of Meta's Llama-3.3-70B-Instruct), specifically designed for complex reasoning and agent tasks.
While achieving SOTA performance in science, math, programming, and agent tasks, the model also triples throughput compared to its predecessor and can run efficiently on a single card.
Llama Nemotron Super v1.5 uses Neural Architecture Search (NAS), allowing the model to strike a good balance between accuracy and efficiency, effectively converting throughput gains into lower operational costs.
As Nvidia's latest open-source large language model, Llama Nemotron Super v1.5 belongs to the Nvidia Nemotron ecosystem, which integrates large language models, training and inference frameworks, optimization tools, and enterprise-grade deployment solutions, aiming to enable high-performance, controllable, and easily scalable generative AI application development.
Related links: https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5 (@量子位)
2. Qwen3 Releases New Open-Source Model Qwen3-30B-A3B-Instruct-2507
Alibaba's Tongyi Qianwen team has released a brand new open-source model, Qwen3-30B-A3B-Instruct-2507.
This model uses a 'non-thinking mode'. 'Non-thinking mode' is an innovative approach by Tongyi in model architecture, aiming to deliver high performance at extremely low computational cost. By activating only the core 3B parameters of the model, you can experience performance far exceeding traditional lightweight models.
The model has the following key improvements:
- Significantly improved general capabilities, including instruction following, logical reasoning, text comprehension, math, science, programming, and tool use;
- Greatly increased long-tail knowledge coverage across multiple languages;
- Better alignment with user preferences in subjective and open-ended tasks, providing more helpful answers and higher-quality text generation;
- Enhanced capabilities in understanding 256K long contexts.
It is now fully open on ModelScope and HuggingFace, and developers can download and use it for free right now.
Related links: https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507 (@ AIGC 面面观、@智深 AI 观)
02 Highlighted Products
1. Microsoft's 'AI Browser' Launches, Supporting Voice and Text Commands and Voice Navigation
Recently, Microsoft launched a new experimental feature in the Edge browser: 'Copilot Mode', which completely transforms Edge from a traditional web browser into an 'AI-native' browser.
Currently, Copilot Mode is still considered an experimental feature and is undergoing testing and iteration. Users can ask questions like in ChatGPT, have it understand webpage content, initiate actions, and even attempt to complete tasks.
According to Microsoft's official blog, the new Copilot Mode brings a brand new tab page, global context browsing across tabs, and Action capabilities that can handle various tasks.
The traditional new tab page is replaced by a minimalist input box. This input box integrates chat, search, and web navigation. After entering a query, compared to the previous Edge which directly searched, Copilot now attempts to predict intent based on already open tabs, allowing users to start tasks faster rather than passively waiting to click.
In addition to summarization and conversation, Copilot Mode adds a new capability called 'Actions', supporting users to issue commands via voice or text for the AI to 'execute operations'.
Currently, Copilot also supports voice navigation, which is indeed more user-friendly for those unfamiliar with computer operations or users with mobility issues. We can directly say to it: 'Help me open a few pages and compare the iPhone 15 and S24 Ultra', and Copilot will execute the operation.
Microsoft also showcased upcoming features. In the future, after obtaining user authorization for browsing history and even login credentials, Copilot will be able to execute more complex 'Agent tasks'.
In the demo video, simply saying 'Help me find a paddleboard rental near the company' allows Copilot not only to find the best option but also to check the weather, complete the booking, and recommend sunscreen and instructional videos.
During the testing period, this feature is completely free to use (it may become a paid subscription in the future), and using Copilot Mode requires users to voluntarily opt-in. (@ APPSO)
2. Google Search AI Mode Launches Multiple New Learning and Exploration Features
Google has launched a series of updates for the AI Mode in its search engine, including support for asking questions about images and PDFs, a new Canvas planning feature, and Search Live real-time conversations with integrated video input.
Google announced a series of new features for AI Mode in its search engine, aimed at helping students, educators, and all curious users. Users can now find a new AI Mode button on the Google homepage on desktop to enable these features.
New features include support for asking questions about images and PDF files on desktop. Users can upload files for complex queries, and the AI will analyze the content and provide answers combined with web information. Support for file types in Google Drive is planned for the future. The new Canvas feature helps users create and organize plans in a dynamic sidebar. The Search Live feature now integrates video input, bringing the advanced capabilities of Project Astra into AI Mode, allowing users to have real-time visual conversations with AI through Lens in the Google app. Additionally, through Lens in the Chrome browser and AI Mode, users can ask questions about anything displayed on the screen.
Features:
- Image/PDF Questions: Upload images or PDFs on desktop for complex queries, AI answers combined with web information;
- Canvas: Create and organize plans, such as exam prep schedules, in a dynamic sidebar;
- Search Live: Integrated video input, real-time visual conversations with AI through Lens;
- Page Questions: Ask questions about the current screen content in the Chrome browser through Lens.
Related links: https://blog.google/products/search/ai-mode-updates-back-to-school/ (@ APPSO)
3. Eigent Released, the First Desktop Multi-Agent Collaboration System
The first desktop multi-agent collaboration system, Eigent, has been released. It consists of a group of AI agents that can collaborate in parallel and has now launched a public beta and is fully open-source. Built by the CAMEL-AI team, Eigent is the world's first desktop Multi-agent Workforce system, built on the open-source projects CAMEL (13k Stars) and OWL (17k Stars).
A desktop multi-agent system named Eigent has been released. Eigent is described as the first multi-agent workforce on the desktop, consisting of a group of AI agents capable of parallel collaboration to complete complex tasks. The system is positioned as a long-term work partner for users, offering fully customizable workers and MCPs.
Eigent has currently launched a public beta for MacOS and Windows and is 100% open-source on GitHub.
- Custom Agents: Users can customize their own agents according to their industry needs, such as a legal agent that understands compliance, a social media editor that captures keywords, a search agent skilled in research, etc.;
- Multi-agent Parallel Execution;
- Human-in-the-loop intervention at any time;
- Flexible access to massive MCP tools: The built-in MCP tool library already has over 200 types, and it also supports uploading commonly used MCP tools to equip agents with more suitable tools;
- 100% Open-source + Local Deployment: Not only free to use, but also allows users to connect their own local models.
Related links: https://github.com/eigent-ai/eigent
Official website link: https://www.eigent.ai/ (@ CAMEL AI)
4. Bragi Partners with OpenAI to Introduce GPT-4.1 Voice Assistant to Third-Party Headphones
On July 29, German audio technology company Bragi announced a partnership with OpenAI to launch a technology solution called ChatAI, designed to allow third-party headphone and audio device manufacturers to easily integrate a generative AI voice assistant powered by GPT-4.1.
This partnership marks an important step for Bragi in the AI field since its pivot from hardware manufacturing to technology licensing in 2019. Through the ChatAI platform, any headphone brand can add advanced AI features to its products, such as instant Q&A, information queries, and voice notes, which users can activate with a button press or wake word.
Core Technology Highlights:
AI Model and Processing Pipeline:
- The core engine uses OpenAI's GPT-4.1 model.
- User voice requests are first pre-processed and formatted by a multimodal local model within the ChatAI application, then securely uploaded to OpenAI, and finally, the generated text response is played back in the headphones via speech synthesis.
Intelligent Recognition and Integration:
- Built-in speaker recognition technology can learn and recognize the user's voice, minimizing false triggers from ambient noise and other people's conversations.
- Collaborating with Apple on MFi certification to achieve more seamless background communication and iOS device experience.
Privacy Protection:
- Bragi commits that its servers do not receive or store any user data. All requests are sent directly from the local device to OpenAI, and local data is not retained after the request is sent.
Platform Extensibility:
- Allows partner brands to build customized and branded experiences on top of ChatAI, and even develop 'Agentic' features, such as proactively adjusting the headphone's equalizer (EQ) settings based on user habits and scenarios.
Release Plan and Market Information:
- Launch Time: The ChatAI application is planned for official launch this summer, but specific headphone products equipped with this technology will be announced by the respective partner brands.
- Supported Languages: Initially supporting English, German, French, Spanish, and Chinese, with plans to expand to more languages in the future.
- Market Outlook: A survey shared by Bragi shows that 89% of consumers are willing to pay an average of $25 more for headphones with AI capabilities. (@ 9to5Mac)
5. China Telecom's First Smart Glasses Released: Equipped with Tianyi Xingchen AI Large Model, Supporting Full-Duplex, Long-Range Pickup, and Other High-End Dialogue Features
On July 29, China Telecom announced the official release of its first AI glasses — TeleAI GLASSES, priced at 1999 RMB. The product is equipped with a high-performance Xingchen large model. It deeply integrates Tongyi Qianwen technology and features practical tools like instant translation and knowledge Q&A.
It adopts an ultra-light half-frame design, weighing 38 grams, with adjustable nose pads and special temple tips.
Equipped with a 12-megapixel ultra-wide-angle camera, supporting up to 1440P 30fps video recording, and featuring a series of advanced algorithms: from electronic image stabilization to portrait optimization, everything is included.
Built-in Tianyi Xingchen AI large model, with features like intelligent object recognition.
Integrated five-microphone array and Qualcomm ADSP platform, coupled with seven core algorithm supports, endowing these glasses with multiple high-end dialogue features like full-duplex and long-range pickup. Also matched with open speakers, etc., with noise reduction algorithms.
Built-in 2GB RAM and 32GB storage, battery life of 7 hours, and the charging case can achieve 10 full charges. (@雨询、@数科雷达站)
6. Alibaba's Ali Xiaohao Stops Renewals, App to be Removed by October 31
IT Home reported on July 29 that Alibaba Communication issued an announcement today (July 29), stating that according to operator requirements, Alibaba Communication will no longer use mobile phone numbers to provide number protection services for individual users.
The announcement stated that the Ali Xiaohao service used by users will stop renewals starting July 29, 2025, and will end service and go offline in batches according to the following sequence, with corresponding numbers being reclaimed. If users have registered internet accounts with the number, they should unbind them promptly. Communication balances will be refunded at the purchase price within 10 working days starting from the day after the number's service ends and goes offline.
IT Home Note: Ali Xiaohao is a virtual mobile number service provided by Alibaba Communication, a subsidiary of Alibaba Group. It was shut down and taken offline on April 29, 2025. If users have registered internet accounts using an Ali Xiaohao number, it is recommended to unbind them promptly to avoid inconvenience caused by number reclamation. (@IT Home, @Jiangnan Metropolis Daily)
03 Opinionated Viewpoints
1. WAIC2025 Concludes: Signed Investment Exceeds 15 Billion
Recently, the three-day 2025 World Artificial Intelligence Conference (WAIC2025) officially concluded in Shanghai.
It is reported that this year's conference gathered over 1,500 top experts from more than 70 countries and regions, including 12 Turing Award and Nobel Prize laureates, over 80 Chinese and foreign academicians, and representatives from multiple international top laboratories. Over 800 enterprises participated, comprehensively showcasing new trends and latest achievements in the development of large models, multimodal AI, and embodied intelligence.
According to a report by Cailian Press, at the closing ceremony on July 28, a batch of major projects were signed centrally, covering fields such as intelligent driving, embodied intelligence, and robotics. A total of 31 projects were signed, with an investment exceeding 15 billion RMB. According to data from the organizers:
- Exhibition area: This year's conference exceeded 70,000 square meters. The main exhibition area for WAIC in 2023 was 50,000 square meters, and in 2024 it exceeded 52,000 square meters;
- Exhibits: This year showcased over 3,000 cutting-edge technologies, including more than 100 'global debut' and 'China debut' products, achieving a 'double increase' in total exhibits and debut products;
- Intelligent robots were also a highlight: In 2023, over 20 robots were showcased; in 2024, a matrix display of 18 humanoid robots was staged; this year, over 60 intelligent robots participated, with a total of over 150 humanoid robots. (@ APPSO)
2. Former Google CEO: Open Source Has Become an Important Feature in AI Development
On July 26, at the World Artificial Intelligence Conference (WAIC), former Google CEO Eric Schmidt engaged in a 'fireside chat' with Harry Shum, Chairman of the Board of Trustees of the Hong Kong University of Science and Technology, on the theme of 'Prospects for Global AI Cooperation'.
Schmidt emphasized that over the past two years, China's AI technology, particularly large models like DeepSeek, MiniMax, and Kimi, has achieved remarkable accomplishments that have attracted worldwide attention. Schmidt pointed out, 'In China, not all of these leading AI models adopt a closed strategy like those in the US.' This has also become an important feature in current AI development.
During the conversation, Schmidt admitted that he personally leans towards supporting open source. He stated that while open-source technology has some potential risks, humanity can manage these risks by setting certain restrictions and adjust them as needed. Regarding this situation, Schmidt believes:
The key question is who the 'editors' of the technology are, where these technologies are applied, and where to set up 'guardrails'. The ideal scenario is that we can train and align these models according to human values.
Additionally, Schmidt believes that collaboration between future super-intelligences is ultimately inevitable. Because as technology develops, humanity will possess a super-intelligent system, and in the future, these systems will have the ability to collaborate and coordinate with each other. For this situation, Schmidt believes 'it is necessary to allow researchers from China and the West to communicate with each other, collaborate and discuss, and reach a consensus on values'. (@ APPSO)
More Voice Agent Learning Notes:
Written at the End:
We welcome more partners to participate in the co-creation of the 'RTE Developer Daily' content. Interested friends, please contact us through the developer community or public account messages, and remember to use the code word 'co-creation'.
For any feedback (including but not limited to content and format), we are extremely grateful and have small surprises in return. For example, what content you hope to see in the daily report; sources, projects, topics, events you recommend yourself; or list a few content channels you like and frequently read; areas where content layout or presentation format can be improved, etc.
Source materials from official media/online news