AI Daily: DouBao 2.0 to be released on Valentine's Day; MiniMAX M2.5 may launch first overseas; Xiaomi opensources its first robot VLA large model

Welcome to the "AI Daily" column! This is your guide to exploring the world of artificial intelligence every day. Every day, we present you with the latest content in the AI field, focusing on developers to help you understand technology trends and innovative AI product applications.

Fresh AI products Click to learn more:https://app.aibase.com/zh

1. Valentine's Day Big News! Volcano Engine's Douba 2.0 Will Be Launched: Video Generation Directly Delivered to "Industrial-Level" Delivery

Volcano Engine, a subsidiary of ByteDance, plans to release multiple technical upgrades on February 14th, focusing on the 2.0 version of the "Douba" series, including the video and audio creation tool Seedance 2.0 and the image creation tool Seedream 5.0 Preview. This upgrade aims to lower the threshold for professional content production, marking an important step for ByteDance in digital creation and enterprise efficiency tools.

AiBase Highlights:

Industrial-level video delivery: Seedance 2.0 reaches industry-leading levels in human-machine interaction and dynamic image stability, supporting full-modal input, with output quality suitable for professional scenarios such as film and advertising.

Real-time retrieval enhancement: Seedream introduces real-time information retrieval for the first time, ensuring creative content stays synchronized with social timeliness and accurately meets immediate creative needs.

Deep intent understanding: The new system includes an encyclopedic knowledge base, significantly enhancing multilingual processing capabilities. Even with vague input, it can achieve high-precision image-text alignment and subject consistency.

2. Going Global First! MiniMAX M2.5 Model Will Be Officially Launched: Overseas Agent Has Started Secret Internal Testing

The overseas internal testing of the MiniMAX M2.5 model marks an important advancement in its global expansion and demonstrates its technological competitiveness in the intelligent agent field.

AiBase Highlights:

🧠 MiniMAX M2.5 model has started internal testing in overseas agent products, showing its ambition for global expansion.

🤖 The model focuses on intelligent agent (Agent) scenarios, improving task planning and complex command execution capabilities.

🚀 As internal testing progresses, M2.5 will accelerate its global launch for developers, enriching the product matrix of domestic large models.

3. Xiaomi Opensources Its First Robot VLA Large Model, Breaking the "Physical Intelligence" Delay Bottleneck

Xiaomi opensources its first robot VLA large model, Xiaomi-Robotics-0, which has 4.7 billion parameters and solves the problem of slow robot actions caused by inference delays in existing VLA models, achieving real-time inference and efficient generalization on consumer-grade graphics cards.

AiBase Highlights:

🧠 Visual Language Brain (VLM) is responsible for parsing human instructions and capturing spatial relationships.

⚙️ Action Expert (Action Expert) generates precise action blocks through flow matching technology.

🚀 Two-phase training improves the model's response agility in real environments.

Details: https://xiaomi-robotics-0.github.io

4. Integrated Mainstream Large Models! Baidu Qianfan Launches AI Coding Subscription Service Coding Plan

Baidu Qianfan launches a new AI coding subscription service, Coding Plan, deeply integrating mainstream code models such as GLM-4.7 and DeepSeek-V3.2, supporting one-click switching, compatible with tools using OpenAI and Anthropic protocols, and offering diverse subscription plans to meet the needs of individual developers and high-frequency users.

AiBase Highlights:

🧠 Deeply integrated with multiple mainstream code models to improve development efficiency.

🔄 Supports one-click model switching, simplifying the development process.

💰 Offers various subscription plans to meet different needs.

5. Zhipu Launches GLM-5: From "Code Generation" to the Agentic Ready Era of "Engineering Construction"

Zhipu launches GLM-5, marking the transition of large models from code generation to the Agentic Ready era of engineering construction. Its performance is comparable to industry leaders, with core breakthroughs including parameter expansion, asynchronous reinforcement learning, and sparse attention mechanisms, while also supporting multi-agent development and adapting to domestic chip platforms.

AiBase Highlights:

🧠 GLM-5 model expands parameters to 744B, with pre-training data increased to 28.5T.

🚀 Uses the "Slime" framework and asynchronous agent reinforcement learning algorithm to enhance long-range interaction capabilities.

💻 Supports multi-agent concurrent completion of the entire development process and adapts to domestic chip platforms.

6. DeepSeek Receives a Version Update: Context Length Jumped to 1M, Trillion-Parameter Flagship Model on the Way

The DeepSeek web and mobile app have received a version update, with context length increased to 1 million Tokens, marking its entry into the global top tier in large document processing. This upgrade increases processing power by nearly 8 times and can handle an entire long novel at once, with wide application scenarios. Meanwhile, the next flagship model, DeepSeek V4, is in the final stages of development, expected to reach a scale of trillions of parameters.

AiBase Highlights:

🚀 Context length increased to 1M Tokens, significantly enhancing processing power.

📚 Can handle entire long novels, suitable for various scenarios.

🧠 Next-generation flagship model DeepSeek V4 is under development, with a scale reaching trillions of parameters.

7. Refuse "Stupid" Glasses! Rokid Glasses Support Access to Private Models Such as DeepSeek/Kimi, You Define Your Own Glasses

The article introduces the new "Custom Intelligent Agent" feature of the Rokid AI glasses, Rokid Glasses, which allows developers to integrate private models into the glasses system, enhancing hardware playability and promoting the transformation of AI wearable devices toward "user-defined" solutions.

AiBase Highlights:

🧠 Wide model compatibility: Developers can access popular models such as DeepSeek R1, Qwen3, and Kimi K2.5 through standardized interfaces.

💻 Native support for open-source frameworks: Supports direct integration with the OpenClaw open-source framework, enhancing logical processing capabilities.

🔒 Security responsibility: Users are responsible for the data security and compliance of custom intelligent agents and must abide by cybersecurity regulations.

8. A Developer's Blessing! Baidu Qianfan Coding Plan Officially Launched: GLM and DeepSeek Combine Strongly, AI Coding Enters the "Subscription Freedom" Era

Baidu Qianfan launched the AI coding subscription service Coding Plan, marking its upgrade from tool integration to a complete service loop in the code development field, aiming to provide developers with a more flexible and efficient intelligent coding experience.

AiBase Highlights:

🧠 Full-scenario coverage: Service capabilities fully cover the entire process of code writing, logic understanding, and system optimization.

🤖 Mainstream model integration: The first official integration includes multiple globally leading code generation and understanding models such as GLM-4.7 and DeepSeek-V3.2.

🔄 One-click seamless switching: Developers can freely switch different model backends in the control panel based on task requirements, without modifying any local environment configurations, greatly enhancing development flexibility.