Summary of Key Points
Recently, the “open models” developed by Chinese AI companies such as DeepSeek, Zhipu, and Kimi have attracted significant attention. Coupled with the “Open Model Alliance” promoted by NVIDIA CEO Jensen Huang, the trend towards openness in the AI industry has evolved from a strategic move by Chinese firms to a global industrial phenomenon. However, it’s important to note that current “open models” are not necessarily “open-source”; they represent a form of “open weights.” The core of this news article is to dissect what these “open weights” actually offer to users.
1. Understanding the Difference Between “Open Models” and “True Open-Source”
Many people mistakenly believe that “open models” are the same as “open-source,” but there’s a significant difference:
- True open-source: This means providing you with a complete set of resources, including the AI algorithm, training data, and code. For example, the early competitor to ChatGPT, Llama 2, was available in an open-source version. However, even open-source models may have certain restrictions.
- Open weights: In this case, the model is already trained, and you are only given the “weight files” that define its behavior (the parameters that determine how it functions). You can use the model directly for tasks like generating text or images, but you may not have access to the training code, and there might be restrictions on commercial use (such as sharing profits or limitations on application areas).
In simple terms, open-source allows you to build your own AI systems, while open weights enable you to use pre-trained models developed by others.
2. What Exactly Are These “OpenWeights” Offering?
Don’t let the term “open” mislead you; they mainly include the following:
1. Model weight files: These digital files contain the parameters that define how the AI model functions. By making these files available, users can use the trained models without spending millions or billions on training processes. Of course, you need a powerful computer to run them.
2. Basic usage rights: Some companies allow users to use the models for personal projects or limited commercial purposes, though there may be restrictions (e.g., fees if monthly revenue exceeds a certain threshold). Companies often offer both free and paid versions, with the paid version providing more stable performance or advanced features.
3. Some auxiliary tools: This might include tutorials on how to deploy the model or simple tools for fine-tuning it (to make it more suitable for specific applications, such as in healthcare). However, the core training code and raw data are usually not provided.
3. Why Is the Trend of “OpenWeights” Popular Now?
Companies prefer to offer open weights rather than full open-source solutions for several reasons:
- Protecting core technology: By only sharing the results, they can keep their training methods and data sources confidential.
- Faster ecosystem development: This encourages more developers to create applications using their models (e.g., AI customer service tools or content generation software), making their models central to an entire ecosystem. For example, NVIDIA’s Open Model Alliance aims to promote the use of its GPUs for model training, thereby driving hardware sales.
- Low-cost monetization: They can attract users with a free basic version and then generate revenue through paid advanced services or licensing fees.
4. What Are the Benefits of OpenWeights for Us?
Both individuals and small businesses can benefit from open weights:
- Small companies/developers: They can use pre-trained models to develop their own products without investing in expensive training processes. For example, a team can customize an AI education tool using Zhipu’s model, saving 90% on development costs.
- Ordinary users: There are now many affordable or free AI tools available based on open weights, making AI more accessible.
- Industry progress: Open weights encourage more people to contribute to AI improvements. For instance, some companies use these models for medical diagnosis aids or agricultural forecasting, which can benefit our daily lives.
5. What Are the Potential Risks?
While open weights offer many benefits, there are also potential drawbacks:
- Commercial restrictions: Some models come with terms requiring profit sharing or limitations on usage (e.g., in military or fake information applications).
- Security risks: If malicious actors use these models for fraudulent purposes, companies may be held responsible, as users are often considered secondary users.
- Ecosystem dominance: Large companies may gain control of the ecosystem, leaving smaller firms limited to developing niche applications and unable to develop their own core technologies.
In summary, open weights represent a step towards greater collaboration in the AI industry. While they offer significant benefits for individuals and small businesses, it’s crucial to understand the associated rules and risks. In the future, both true open-source and open-weight models are likely to coexist, making AI technology more widespread and practical.