Summary of the Key Points
The K3 large model, launched by Chinese company “MoonZhiDanMian” (with 2.8 trillion parameters, ranking third in the world), has been made available for free download via open source. However, ordinary individuals or small and medium-sized enterprises (SMEs) are simply unable to deploy it on their own due to the exorbitant costs associated with the required hardware and electricity. While open source may seem “free,” it has raised the barrier from a legal permission issue to a practical one of having the financial means to build the necessary infrastructure. Hardware manufacturers like NVIDIA, on the other hand, stand to profit significantly from the widespread adoption of these models, as enterprises prefer to use them through API calls rather than building their own installations.
1. The “Free” Nature of Open Source Large Models: Sounds Good, but Not So Easy to Access?
Many people assume that “open source” means it’s free to use without any restrictions. However, the open source of K3 is different from familiar software like Linux or Firefox:
- What’s open-sourced isn’t just the code: While typical open-source software provides the source code that can be modified, K3 offers a pre-trained model with 2.8 trillion numerical matrices, packaged in 1.4TB of data. You can fine-tune it, but you won’t understand the meaning of each number—it’s like having a Michelin cookbook without access to a top-tier kitchen.
- The barriers to use are incredibly high: K3 is a “Mixed Expert Model” (MoE), similar to a hospital with 896 departments; only 16 departments need to be active at any given time, but all must be ready to respond immediately. To run it smoothly, the official recommendations include at least 64 high-performance GPUs (costing approximately 17 million RMB) and specialized power infrastructure (each GPU consumes 45 kilowatts of electricity, four times what a typical household uses). Even if the model is free, your home’s electrical system wouldn’t be capable of supporting it.
2. The Barrier from “Legal” to “Physical”
The barrier for closed-source models is financial authorization; for open-source models, it’s the ability to build the necessary infrastructure:
- Closed-source: Not using them is illegal: For example, using OpenAI’s GPT without paying is considered a copyright infringement.
- Open-source: Free to use, but not affordable: K3 is free to download, but running it requires millions in hardware costs, plus renting data centers and hiring maintenance staff, with monthly electricity expenses being astronomical. Traditional software is expensive to develop but free to replicate and run; large models, however, are expensive to develop, free to replicate, but extremely costly to operate. The “free” aspect only applies to replication; the operation still incurs significant costs.
3. API Calls vs. Self-Deployment: Which Is More Cost-Effective?
For enterprises, the choice is clear:
- API calls: A small investment can achieve significant results. Using K3’s API costs only $15 per million tokens produced (three times cheaper than OpenAI), which might amount to a few thousand to tens of thousands of RMB per month—a reasonable expense.
- Self-deployment: It’s a costly endeavor. Even with extreme efficiency, you would need 16 high-performance GPUs (costing millions) and still have to account for electricity, data center expenses, and maintenance staff. Unless you have specific compliance requirements (such as preventing data from leaving the company’s internal network), using APIs is the more cost-effective option.
- Single-machine deployment? Out of the question: The previous generation of K2 (with 1 trillion parameters) could barely be run on a high-end Mac, but K3’s requirement of 2.8 trillion parameters makes single-machine deployment impossible; it’s clearly designed for data centers or large enterprises.
4. Why Do Giants Like NVIDIA Support Open Source?
NVIDIA CEO Jensen Huang’s support for open-source models might seem unusual but is actually strategic:
- Open source models create a business opportunity: With models available for free, users need to purchase more GPUs (NVIDIA’s core product). The more popular K3 becomes, the more GPUs are needed, and the more NVIDIA profits. It’s similar to the gold rush, where those who sell tools always profit.
- The logic behind Silicon Valley giants’ support: Companies like Microsoft and Meta sign open-source initiatives because they recognize that open source can challenge the dominance of closed-source models (like OpenAI). They either sell hardware (NVIDIA) or cloud services (Microsoft), and more open-source models mean greater demand for their products.
5. What Should Ordinary People Think About Open Source Large Models?
Next time you see news about an open-source model, don’t just assume it’s free. Ask yourself three questions:
1. How many GPUs are required to run it? High-end GPUs cost tens of thousands each, so dozens would amount to millions.
2. Is the electricity supply sufficient? A typical household’s power system can’t handle such demands.
3. Are the maintenance costs high? It requires a professional team to manage.
For ordinary people, using APIs or existing services (like Kimi APP) is sufficient. For business owners, don’t be misled by the “free” label; carefully consider the hardware costs before deciding whether to deploy the model.
In conclusion: The so-called “free” nature of open-source large models is actually aimed at those who have the resources to build the necessary infrastructure. For most people, using APIs offers the best value for money.