Articles

This article presents six significant open-source artificial intelligence projects. These include ZTE's Embodied Intelligence brain model EmbodiedBrain1.0, which employs SFT+RL dual training to tackle challenges in Agent Alignment, Real-time Performance, and real-world evaluation within Embodied AI; Nous Research's Mixture-of-Experts Reasoning model Hermes 4.3, demonstrating superior performance in mathematics, code, and related domains; the Autoregressive Flow Model STARFlow for text-to-image and video generation; the CUDA-L2 system, integrating LLMs and RL for automated CUDA Kernel optimization; the ReCode project, enabling adaptive decision-making through recursive code generation; and the AN1 Meaning Engine, based on PyTorch, for efficient semantic extraction. The article aims to equip technology practitioners with a concise understanding and facilitate exploration of the latest open-source AI advancements.

This issue of 'Interesting Weekly' features a selection of tech articles covering the evolution of Function Calling, MCP, A2A, and Skills in the AI domain; the optimal SEO learning roadmap and book recommendations; B2B product scenario analysis and solution design methods; practical experience in AI complex task processing shared by Taobao engineers; eight mainstream methods for strategy product weight design; and the thirty-year technological evolution of recommendation systems. In addition, the weekly introduces various AI tools, including Nano Banana Pro, an image generation model based on Gemini 3 Pro; Next AI Drawio, an AI-powered chart creation tool; AlphaBot, a smart stock analysis assistant; WinDog, a quantitative trading system; and AI-Stock-Master, an AI stock trend analysis platform. The articles cover a wide range of topics, aiming to provide readers with a curated summary of daily reading and thoughts for the week.

This article captures an in-depth dialogue between Luo Yonghao and Zhang Peng at the GeekPark IF 2026 Conference. Luo Yonghao candidly revealed the less glamorous origin of his podcast: to address a funding crunch for his new AI hardware project, leveraging his influence to "fund his ventures." He detailed the podcast's operational logic, from topic selection and interviews to commercialization, highlighting his steadfast principle of not accepting paid appearances to prevent content from being compromised by capital. During the conversation, he expressed both admiration and concern for today's young entrepreneurs, attributing their abilities to a high-quality growth environment and recognizing the unprecedented opportunities presented by the AI revolution. Simultaneously, Luo Yonghao shared his enduring passion for products and his dedication to "AI phones," outlining his strategy for entrepreneurship in a market dominated by giants and encouraging young people to "not be afraid."

This article details the Central Reallocation Hashing (CRH) method, jointly proposed by Beijing University of Posts and Telecommunications, Beihang University, and China Telecom, to address the limitations of deep hashing methods in large-scale image retrieval. Traditional hash center-based methods often overlook inter-class semantic relationships, while two-stage methods introduce complexity and bias despite attempting to incorporate semantics. CRH innovatively uses an end-to-end framework, iteratively updating hash centers through a dynamic reallocation mechanism during hash function training, eliminating the need for complex pre-training or offline optimization. The method centers on constructing a hash codebook, optimizing the hash function using margin cross-entropy loss and quantization loss, and periodically reallocating hash centers based on the average distance between sample hash codes and codebook centers. Furthermore, the multi-head codebook design expands the semantic representation ability of the codebook from M to M^H^, enhancing semantic expression. Experimental results demonstrate that CRH outperforms existing state-of-the-art methods on benchmark datasets like Stanford Cars, NABirds, and MS COCO, particularly in handling fine-grained semantic relationships. Ablation studies confirm the effectiveness of central reallocation and multi-head mechanisms and demonstrate high robustness. This research offers a novel approach to deep hashing learning. The code is open-sourced and accepted to AAAI 2026.

The author developed a pure Dart image editor for Flutter due to limitations in UI consistency, desktop support, scalability, and performance of the existing `image_cropper` plugin. This editor utilizes `CustomPaint` and matrix transformations to achieve pixel-perfect lossless cropping, built-in zoom compression, a cross-platform unified experience, 360° rotation, text addition and editing, and undo/redo functions. The article details the editor's layered architecture design (View, Controller, Model, Handlers/Managers), explains the rendering mechanism of `CustomPaint`, the principle of high-fidelity cropping based on matrix transformations, and the implementation details of text and history record management. In addition, it focuses on how to achieve high-quality client-side image compression through configurable `ImageCompressionConfig` and `drawImageRect`, effectively reducing the server load. The article not only provides a detailed technical implementation plan but also includes links and usage examples of the open-source library, which has extremely high reference and practical value for Flutter developers.

The article cites the latest research from Nature and the views of Google's Ilya Sutskever, subverting the traditional understanding based on Moore's Law regarding intelligence development. It points out that in the past few decades, although chip speeds have peaked, AI intelligence has leaped forward. The core driving force is not single-point acceleration, but the structural merging and parallel collaboration of computing units. The article likens this to the evolution of biological intelligence. It emphasizes that the essence of intelligence is improving prediction ability, achieved through large-scale collaboration. The rise of AI follows this bio-inspired evolution model, that is, through large-scale parallel computing and multi-module collaboration, new intelligence emerges that exceeds individual capabilities. The article further explores that future intelligence will be a distributed, symbiotic collaboration network of human-machine collaboration, rather than a single entity.

This article details the NVIDIA Run:ai Model Streamer's new native integration with Google Cloud Storage, directly addressing the critical "cold start" problem for large language model (LLM) inference on Google Kubernetes Engine (GKE). As LLMs grow, the time taken to load models into GPU memory becomes a major bottleneck, leading to idle GPUs and delayed service scaling. The Model Streamer resolves this by streaming model tensors concurrently to GPU memory, cutting loading times from minutes to seconds. For multi-GPU setups, its distributed streaming capability efficiently coordinates loading across GPUs using NVIDIA NVLink. The integration offers simplicity, performance, and enhanced security, enabling vLLM users to activate it with a single command-line flag and leverage GKE's Workload Identity for simplified credentials. Furthermore, combining the Model Streamer with Cloud Storage Anywhere Cache provides additional server-side acceleration, creating an exceptionally performant model-loading system. This collaboration aims to foster faster, more resilient, and flexible MLOps pipelines on GKE.

This article details Rokid's outstanding performance and key achievements in the smart glasses industry in 2025. Facing explosive growth in the smart glasses market, Rokid has meticulously refined its core product capabilities, distilling features like photography, translation, teleprompting, navigation, and payment into compact, lightweight glasses. It has also innovated with magnetic charging, optimized a dual teleprompter mode, addressed light leakage issues, upgraded AI quick commands, and enhanced the payment experience. Furthermore, Rokid has paid close attention to the special needs of users with myopia and the visually impaired community. Its products have achieved skyrocketing sales in both domestic and international markets, setting multiple records. Rokid has also built a robust ecosystem through collaborations with platforms like Alipay and Amap. Beyond its success in the consumer market, Rokid's products have been adopted in professional, high-intensity scenarios by China Central Television (CCTV), the Haier Group CEO, and others, transitioning smart glasses from 'experimental devices' to 'trusted devices' and cementing Rokid's role as an industry leader.

This article provides a detailed overview of six cutting-edge AI open-source projects. Key among them is Devstral Small 2, a lightweight Agent-type Large Language Model (LLM) tailored for software engineering tasks and capable of running on local devices. Also featured are several powerful frameworks and tools: the GLM-ASR-Nano-2512 speech recognition model, which excels in Mandarin dialect recognition and low-volume robustness; the Fairy2i-W2 universal framework, achieving extremely low-bit quantization while maintaining LLaMA-2 7B performance; and siiRL, a distributed reinforcement learning framework designed to mitigate scalability bottlenecks in LLM post-training. Additionally, the Open-AutoGLM mobile intelligent assistant framework and Mistral Vibe command-line coding assistant are introduced, both showcasing advanced technology and significant application potential in their respective domains. It's worth noting that the article focuses on project summaries, offering less in-depth analysis of core technical principles and innovations, making it ideal for a quick grasp of current technological trends.

Presented as an "Open Source Daily," this article recommends five of the latest GitHub open-source projects. First among them is SharpKeys, a Windows key remapping tool that enables users to customize keyboard functions via a graphical interface, eliminating the need for manual registry modifications. Second is OWASP's WebGoat, a deliberately insecure web application designed to educate users on cybersecurity knowledge and penetration testing techniques through practical exercises. Following this is the Ministry of Justice's modernisation-platform project, which utilizes Terraform for defining and managing cloud infrastructure. The fourth project is rasbt/scipy2023-deeplearning, offering deep learning workshop materials primarily based on PyTorch. Lastly, category-labs/monad is a high-performance EVM-compatible blockchain node execution component responsible for transaction processing and blockchain state management. The article as a whole is rich in content, demonstrating the application and development of open-source technology across various fields.