AI Weekly: DeepSeek Multimodal Model, Xiaohongshu Leadership Change, Unitree Humanoid Robot

Introduction

This week saw a flurry of activity across the AI and tech landscape, spanning breakthroughs in foundational model technology, consumer-grade robotics products hitting the market, and major organizational reshuffles at internet giants. DeepSeek scored another win in the multimodal space, Xiaohongshu welcomed a new president, and Unitree Robotics pushed humanoid robot pricing to new lows. Here's a detailed breakdown.

DeepSeek Releases Multimodal Model: Redefining Reasoning with "Visual Primitives"

DeepSeek recently officially released its multimodal model on GitHub, along with a full technical report, drawing widespread attention from the industry.

The technical report notes that despite significant advances in multimodal large language models (MLLMs) in recent years, mainstream chain-of-thought (CoT) reasoning paradigms remain largely confined to the linguistic domain. While recent studies have attempted to bridge the "perception gap" through techniques such as high-resolution cropping, the DeepSeek team argues that this overlooks a more fundamental bottleneck — the "referencing gap." The inherent ambiguity of natural language often fails to provide precise, unambiguous guidance for complex spatial layouts, leading to logical breakdowns in tasks requiring rigorous referencing.

To address this issue, DeepSeek proposed an innovative reasoning framework in its technical report: "Thinking with Visual Primitives." This framework elevates spatial markers such as points and bounding boxes to "fundamental units of thought," directly integrating them into the model's reasoning process. In other words, the model can simultaneously "refer to" specific visual locations while "reasoning," effectively anchoring its cognitive trajectory to the physical coordinates of an image.

Notably, DeepSeek emphasized that its framework is built on a highly optimized architecture with exceptional visual token efficiency. Despite a relatively compact model size and a significantly lower image token budget, the multimodal model can already compete with leading models such as GPT-5.4 and Claude-Sonnet on challenging counting and spatial reasoning benchmarks. This once again demonstrates DeepSeek's technical prowess in its "doing more with less" approach.

From a technology trend perspective, this release by DeepSeek signals that multimodal reasoning is transitioning from "understanding images" to "thinking with images." Visual information is no longer merely an auxiliary input but is deeply integrated into the model's reasoning chain.

Xiaohongshu Announces Organizational Restructuring: Conan Appointed as President

On the internet industry front, Xiaohongshu recently announced a major organizational restructuring, officially appointing Conan as the company's president.

As a core member of Xiaohongshu's management team, Conan has been deeply involved in the platform's community operations and commercialization efforts for years. The appointment as president is widely seen as a critical step for Xiaohongshu in accelerating its commercialization and strengthening organizational efficiency. Against the broader backdrop of major internet platforms ramping up AI capabilities and exploring the integration of content and e-commerce, this personnel change has also sparked industry speculation about Xiaohongshu's AI strategy going forward.

In recent years, Xiaohongshu has become increasingly active in its AI initiatives, iterating on features such as AI search, intelligent recommendations, and AI-assisted content creation. The new president's appointment may inject fresh momentum into Xiaohongshu's efforts to integrate AI with its community ecosystem.

Unitree Robotics Unveils Dual-Arm Humanoid Robot: Starting at 26,900 RMB

In the embodied intelligence space, Unitree Robotics launched a new dual-arm humanoid robot with a starting price of just 26,900 RMB (approximately $3,700), once again setting a new price threshold for humanoid robots.

Unitree previously gained global recognition with its quadruped robot products, and this official foray into dual-arm humanoid robotics marks an extension of its product line toward more advanced general-purpose robots. The starting price of 26,900 RMB is extremely competitive for a dual-arm humanoid robot and is expected to help push humanoid robots beyond laboratories and showrooms into broader application scenarios, including education and research, lightweight industrial assistance, and the personal consumer market.

It's worth noting that the humanoid robot sector has been heating up continuously. From Tesla's Optimus to overseas players like Figure and 1X, as well as domestic companies such as Unitree and Agibot, various players are accelerating product iteration and mass production. Unitree's aggressive pricing strategy could reshape the competitive landscape of the industry.

Outlook and Reflections

Although this week's three stories span different sub-sectors, they reflect several core themes in the current AI industry:

• Deepening model capabilities: DeepSeek's multimodal breakthrough shows that model reasoning is evolving from purely linguistic dimensions toward multimodal and spatial awareness. An "efficiency-first" technical approach is challenging the traditional "scale-above-all" paradigm.
• Organizational structures adapting to the AI era: Behind Xiaohongshu's personnel changes is a microcosm of how internet companies are reassessing strategic priorities and realigning organizational structures in the AI wave.
• Hardware democratization accelerating: Unitree's pricing strategy signals that humanoid robots are moving from "proof of concept" to "accessible at scale." Rapidly declining hardware costs will pave the way for the widespread adoption of embodied intelligence.

The evolution of AI technology is simultaneously driving multidimensional transformation across software and hardware, models and products, technology and organizations. The second half of 2025 promises to be one to watch.