AI Daily: ByteDance Launches Seed Prover 1.5; MiniMax M2.1 Open Source; Tongyi Open Sources Speech Interaction Large Model Fun-Audio-Chat-8B

Welcome to the "AI Daily" section! 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 learn about innovative AI product applications.

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

1. ByteDance Launches Seed Prover 1.5: New Advances in Formal Mathematical Reasoning

Seed Prover 1.5, launched by ByteDance's Seed team, has made significant breakthroughs in the field of formal mathematical reasoning. It significantly improves reasoning capabilities and efficiency through Agentic reinforcement learning. The model performed outstandingly in the IMO2025 and Putnam Mathematics Competition, demonstrating its strong performance.

AiBase Summary:

🌱 Seed Prover 1.5 adopts a new Agentic Prover architecture, combining the advantages of natural language reasoning and formal proofs.

💡 The model can flexibly call the Mathlib math library and Python code for verification, enhancing its ability to solve complex problems.

🚀 Seed Prover 1.5 demonstrated excellent performance in the IMO2025 and Putnam Mathematics Competition, achieving silver and gold medal standards.

More details: https://arxiv.org/abs/2512.17260

2. Alibaba Tongyi Open Sources Speech Interaction Large Model Fun-Audio-Chat-8B! Ultra-low latency, able to read emotions

The Alibaba Tongyi Lab officially open-sourced the new end-to-end speech interaction large model Fun-Audio-Chat-8B. This model focuses on ultra-low latency and natural, smooth speech interaction, marking a new stage in open-source speech AI. It not only understands user speech in real-time but also has strong emotional perception capabilities, with performance comparable to closed-source giants GPT-4o Audio and Gemini 2.5 Pro.

AiBase Summary:

🧠 Ultra-low latency and efficient design: Using an innovative dual-resolution architecture, it saves nearly 50% of GPU computing resources and significantly improves response speed.

😊 Empathetic emotional understanding: The model can perceive user emotions from tone, speaking speed, and pauses, and provide considerate and empathetic responses.

🛠 Strong voice function calling: Supports Voice Function Calling, allowing users to perform complex tasks with natural voice commands.

More details: https://funaudiollm.github.io/funaudiochat/

3. MiniMax M2.1 Makes a Stunning Open Source Debut! 10 Billion Activated Parameters Coding Model Tops SOTA, Outperforming Gemini3Pro and Claude 4.5 in Multilingual Programming

MiniMax M2.1 open source large model performs well in multilingual programming and real engineering tasks, surpassing closed-source models such as Google Gemini3Pro and Anthropic Claude4.5 Sonnet, marking a new phase where open source coding models outperform closed-source ones in performance.

AiBase Summary:

🧠 Multilingual programming SOTA: M2.1 reaches the highest level among open-source models in mainstream languages like Python, JavaScript, and Java.

⚙️ Stronger in real engineering tasks: In SWE-Multilingual, M2.1's code repair accuracy and end-to-end task completion rate significantly exceed those of Gemini3Pro and Claude4.5 Sonnet.

🧩 Intelligent agent collaboration optimization: M2.1 excels in tool calling and API integration, providing a strong foundation for building reliable AI developer agents.

More details: https://www.minimax.io/news/minimax-m21

4. Zhihu's 2025 AI Product List Revealed, Doubao Ranks First

Zhihu released its 2025 AI product list, with Doubao ranking first, followed closely by DeepSeek, Gemini, and Tongyi Qianwen, while Zhihu Direct Answer also entered the top five, showing user recognition and expectations for AI products.

AiBase Summary:

🧠 Doubao ranked first on Zhihu's 2025 AI product list, demonstrating its strong market competitiveness.

🤖 DeepSeek, Gemini, and Tongyi Qianwen followed closely, showing their popularity and practicality.

🌐 Zhihu Direct Answer entered the top five, indicating user recognition and expectation of the platform's AI applications.

5. Anthropic Launches a Brand New Open Source Agent Skills Knowledge Base, Enhancing Claude Model Capabilities

Anthropic launched a brand new open source Agent Skills knowledge base, providing practical demonstration skills and best practices for the Claude model's skill system to help developers better utilize the model for various applications. The knowledge base contains 16 public example skills covering multiple fields such as document processing, creative design, development techniques, and enterprise communication. Developers can directly use these skills through the Claude Code plugin marketplace or on Claude.ai (paid users), and they can also create custom skills based on the provided specifications and templates. The Agent Skills functionality allows users to define a series of instructions in the form of packageable and reusable folders, which include Markdown documentation files, scripts, and other resources to guide the model to efficiently complete specific tasks. This feature is now available to all paid users and supports Claude.ai, Claude Code, and its API.

AiBase Summary:

🌟 Anthropic open-sourced the Agent Skills knowledge base, containing 16 demonstration skills covering multiple fields.

🛠 Developers can directly use or customize skills to enhance the application capabilities of the Claude model.

⚙️ Skills functionality enables on-demand loading, reducing Token consumption and improving efficiency.

More details: https://github.com/anthropics/skills

6. Lima v2.0 Released: Evolving from a Container Tool into a "Shield" for Secure AI Workflows

The release of Lima v2.0 marks its important transformation from a container tool to a key security protection for AI workflows. Through innovative virtualization isolation technology, it provides developers with a more efficient and secure AI development environment.

AiBase Summary:

🛡️ Lima v2.0 builds a solid "sandbox" mechanism for AI development, preventing AI agents from accessing host sensitive files or executing high-risk commands.

⚙️ The new version introduces an experimental plugin architecture and GPU acceleration features, improving performance and scalability.

🔒 Integrated Model Context Protocol (MCP) tools, providing a safer local file reading and writing solution.

7. iFlytek Spark Large Model First to Obtain National Security Standards Certification, AI Security Takes a New Step Forward!

iFlytek's Spark large model successfully passed the national security standard pilot test, marking important progress in the field of generative AI security. The model not only participated in the standard formulation work but also established a comprehensive security protection platform and developed a new generation of implicit watermark technology to improve the credibility and security of AI-generated content.

AiBase Summary:

🧠 iFlytek's Spark large model passed the national security standard verification, becoming one of the first organizations to obtain the "Artificial Intelligence Security National Standard Conformity Self-Assessment Certificate".

🛡️ The Spark large model has established a comprehensive security protection platform, covering dozens of technologies, providing "out-of-the-box" protection for multi-language and multi-modal scenarios.

🔐 iFlytek has developed a new generation of implicit watermark technology for generated text, providing a unique digital identity identification for AI-generated content, enhancing credibility.

8. ChatGPT Launches "Year in Review"! Your AI Personality, Annual Poetry, and Creative Medals Are Here, But You Need to Enable Permissions First

OpenAI launched the 'Your Year with ChatGPT' feature, generating personalized annual reviews for users, including interaction data, AI personality medals, and custom poetry. The feature emphasizes privacy protection and is only available to a select group of users.

AiBase Summary:

Generate personalized annual reviews, including interaction data and AI personality medals.

Provide customized poetry and AI paintings to form a unique "annual mental snapshot".

Emphasize privacy first, available only to specific regions and users, and requires manual permission activation.