Tencent Open Sources Hunyuan-A13B: An AI Model with Small Size and Great Intelligence

AIbase基地

Published in AI News · 3 minute read · Jul 9, 2025

Hunyuan-A13B is a newly open-sourced large language model by Tencent. With an innovative design concept, it achieves strong performance with a relatively small number of active parameters, making it especially suitable for resource-constrained environments.

This model adopts a fine-grained MoE (Mixture-of-Experts) architecture, featuring 13 billion active parameters, but the total parameter count reaches up to 800 billion. This design allows it to maintain efficiency and scalability while providing cutting-edge reasoning capabilities and support for general applications.

The core features of Hunyuan-A13B include:

Hybrid reasoning mode supporting fast and slow thinking: This unique reasoning mechanism allows the model to adjust its depth of thinking according to task requirements, improving the efficiency of handling complex problems.
Native 256K ultra-long context understanding capability: This means the model can process extremely long text inputs, making it perform well in tasks that require a lot of background information.
Outstanding performance in agent-related tasks: Hunyuan-A13B demonstrates strong capabilities when performing various agent (Agent)-related tasks.

To achieve efficient reasoning, Hunyuan-A13B adopts Grouped Query Attention (GQA) technology and supports multiple quantization formats. Currently, the model has open-sourced pre-trained, instruction-tuned, FP8, and INT4 quantized versions, making it convenient for developers to use.

In various benchmark tests, Hunyuan-A13B has shown strong competitiveness, especially in mathematics, science, coding, reasoning, and agent domains.

Tencent provides comprehensive support for developers, including detailed guides for interaction and model training using Hugging Face Transformers. Additionally, for model deployment, Hunyuan-A13B offers support through TensorRT-LLM, vLLM, and SGLang, and provides pre-built Docker images and quantized model deployment solutions, greatly simplifying the deployment process.

The open-sourcing of Hunyuan-A13B undoubtedly opens new possibilities for the application of large models in resource-constrained environments and brings new innovation power to the AI community.

Open-source address: https://huggingface.co/tencent/Hunyuan-A13B-Instruct

github: https://github.com/Tencent-Hunyuan/Hunyuan-A13B?tab=readme-ov-file

Alibaba Ovis-U1 Launches with a Bang: A Multi-Modal AI All-in-One, Open Source Empowers Global Developers

On June 29, 2025, the Alibaba International AI Team officially released the new multi-modal large model **Ovis-U1**, marking another major breakthrough in the field of multi-modal artificial intelligence. As the latest masterpiece of the Ovis series, Ovis-U1 integrates multi-modal understanding, image generation, and image editing functions, demonstrating powerful cross-modal processing capabilities, providing new possibilities for developers, researchers, and industry applications. This is a detailed report on Ovis-U1 by AIbase. Ovis-U1

Runway AI Launches Its New Game World: A Large Interactive Text Adventure

Recently, AI technology leader Runway announced the upcoming launch of its new generative AI platform, "Game Worlds." This innovative product marks Runway's successful expansion from the film industry into the gaming sector, offering creators and players a brand-new interactive experience. "Game Worlds": An AI-Driven Interactive Text Adventure. The Runway Game Worlds platform is built on generative AI, allowing users to create and experience text-based adventure games with simple text input. Compared to traditional...

ChatGPT Guides Confused Users to Contact Journalists, Revealing the Impact of AI on User Behavior

Recently, journalist Kashmir Hill from The New York Times exposed a concerning phenomenon: ChatGPT has begun actively guiding users who are caught in conspiracy theories or psychological distress to contact her directly via email. In conversations with users, ChatGPT described Hill as 'empathetic' and 'grounded in reality,' and mentioned that she has conducted in-depth research on artificial intelligence, which may provide understanding and support to these users. Hill mentioned that one of her past contacts was a Manhattan accountant who firmly believed

Pervasive Fake Videos: AI-Generated Content Earns Millions of Views Behind the Diddy Trial

Recently, a wave of fake videos about the trial of "Sean Diddy" Combs emerged on the social media platform YouTube. These videos use AI-generated images and audio accompanied by false information, attracting tens of millions of views. According to an investigation, 26 related channels earned nearly 70 million views from nearly 900 Diddy videos in the past 12 months. The strategies used by these channels are generally similar, often using attention-grabbing titles and AI-generated thumbnails, which lead to

Surprising Similarities Between Large Language Model Search Optimization and Traditional SEO Strategies

Recently, ERGO Innovation Lab and ECODYNAMICS conducted a study focusing on how insurance-related content is displayed in AI-driven search. The research analyzed over 33,000 AI search results and 600 websites, exploring the preferences of large language models (LLMs) such as ChatGPT when processing this content. The study found that LLMs tend to prioritize content that is easy to read, well-structured, and trustworthy, which closely aligns with traditional SEO strategies.