Artificial intelligence is increasingly divided between two different approaches to developing and distributing AI models: closed AI systems and open-weight models.
Companies such as OpenAI, Google, and Anthropic primarily develop proprietary AI systems, where the underlying model weights and much of the development process are kept private. Users typically access these models through applications, APIs, or cloud platforms.
Open-weight models take a different approach. Their trained model weights are made available for others to download and run, allowing developers and organisations to operate the AI on their own hardware or cloud infrastructure.
The distinction matters because it changes how businesses use AI. Instead of sending every request to a model provider’s servers, an organisation can host an open-weight model itself, customise it for particular workloads, and potentially keep sensitive information within its own infrastructure.
The debate also has a growing geopolitical dimension. Chinese companies have invested heavily in openly distributed AI models, while Meta has pursued a similar strategy in the United States. Meanwhile, OpenAI, Google, and Anthropic continue to focus largely on proprietary models.
So what exactly are open-weight models, how do they work, and why are they becoming so important?
What Are Open-Weight AI Models?
An open-weight AI model is one whose trained weights are publicly available.
Model weights are the numerical parameters an AI system learns during training. They encode patterns and relationships that allow the model to process inputs and generate outputs.
When developers receive access to those weights, they can download the model and run it using their own infrastructure, provided the model’s license permits their intended use.
However, downloading a model does not necessarily mean that it will run on an ordinary laptop or gaming PC.
Large AI models can require substantial amounts of:
- GPU memory
- System RAM
- Processing power
- Storage
- Electricity
Smaller models and compressed versions can have much lower hardware requirements, making local AI more accessible.
Several major companies have released open-weight models, including Meta’s Llama family, Alibaba’s Qwen models, Moonshot AI’s Kimi models, and DeepSeek’s models.
How Do Open-Weight Models Work?
With a traditional cloud-based AI service, a user sends a prompt to the company’s infrastructure. The company’s servers process the request and return the response.
An open-weight model changes that setup.
The organization can download the model and provide the hardware needed to run it. That hardware could be a workstation, a dedicated AI server, an on-premises data center, or a cloud environment controlled by the organization.
This provides greater control over where AI processing takes place.
For businesses handling sensitive information, that can be particularly important. Depending on the deployment and security configuration, data can remain inside the organization’s infrastructure rather than being sent to an external AI provider.
Open-weight models can also be adapted for specific applications. Developers may be able to fine-tune or otherwise customize a model for a particular industry, workflow, language, or task, subject to the model’s license.
The Hardware Challenge
The biggest limitation is often hardware.
Large models can require substantial GPU memory and computing resources. Running a model locally can therefore be considerably more expensive or technically demanding than simply accessing an AI service through an API.
This is where smaller models and quantization become important.
Quantization reduces the numerical precision used to represent model parameters, which can significantly reduce memory requirements. Smaller models can also be designed to deliver useful performance while requiring far less hardware.
What Is Model Distillation?
Another approach is model distillation.
In distillation, a smaller model is trained using information generated by a larger or more capable model. The goal is to create a compact system that retains as much of the larger model’s useful behavior as possible while requiring fewer computing resources.
Distillation does not guarantee that the smaller model will perform identically to its larger counterpart. However, it can make advanced AI capabilities much easier to deploy on less powerful hardware.
Open Weight vs Open Source: What’s the Difference?
The terms open weight and open source are often used interchangeably, but they describe different things.
An open-weight model provides access to its trained parameters. That allows users to download and run the model, subject to its license.
An open-source AI project generally goes further by making relevant source code available under terms that allow people to inspect, modify, and redistribute it.
The training data is another important distinction.
A model can have publicly available weights while its:
- Training data remains private
- Training code remains unavailable
- Data-cleaning processes remain private
- Development infrastructure remains proprietary
- Some components remain subject to restrictive licensing
For this reason, an open-weight model is not automatically a fully open-source AI system.
The exact definition can also depend on the license and which components of the model have been released.
Why Are Companies Releasing Open-Weight Models?
Making model weights available may appear to give away valuable technology, but companies can have several strategic reasons for doing it.
Open distribution can help a model build a large developer community. Developers can experiment with it, create applications, fine-tune it, and integrate it into products.
A larger ecosystem can increase the visibility and adoption of the underlying technology.
There can also be commercial benefits.
A company may release model weights while generating revenue from:
- Cloud hosting
- Enterprise services
- AI infrastructure
- Technical support
- Consulting
- Managed deployments
- Hardware and ecosystem partnerships
The model itself can therefore serve to attract developers and customers rather than be the company’s sole source of revenue.
Why Is China Investing in Open-Weight AI?
Chinese AI companies have become major players in the open-weight model ecosystem.
Companies such as Alibaba, Moonshot AI, and DeepSeek have released models that have attracted attention from developers worldwide.
There are several reasons why this strategy can be valuable.
Reducing Dependence on Foreign AI Platforms
Open-weight models can help companies and governments reduce their dependence on foreign AI providers.
Instead of relying entirely on an overseas API, an organization can deploy a model on infrastructure it controls, assuming the model’s license and technical requirements allow it.
This can be particularly important for governments, financial institutions, defense organizations, and other businesses that have strict requirements for data handling and infrastructure control.
Working Around Hardware Constraints
US export controls on advanced semiconductor technology to China have added another layer to the country’s AI strategy.
While hardware remains a critical part of AI development, efficient models can reduce the computing power required for certain workloads.
Open-weight models can also be optimised for different hardware configurations, giving organisations more flexibility in how they deploy AI systems.
Building Global Influence
There is another strategic advantage: adoption.
When developers around the world download and build applications around a model, that model can become part of a broader technology ecosystem.
This gives the company behind the model influence beyond its own applications.
A widely adopted model can lead to third-party tools, integrations, developer libraries, and commercial products that reinforce its position in the AI market.
Why Are Open-Weight Models a Challenge for US AI Companies?
The rise of open-weight models creates competitive pressure for companies whose primary AI products are proprietary.
Companies such as OpenAI and Anthropic spend enormous amounts on model training, infrastructure, research, and engineering. Their business models largely depend on controlling access to their models through products and APIs.
An open-weight alternative changes the economics for some customers.
Instead of paying an AI provider for every API request, a company could purchase or allocate its own hardware and run an open-weight model internally.
That does not make AI free.
Organizations still need to pay for:
- GPUs and servers
- Electricity
- Cooling
- Storage
- Network infrastructure
- Engineers
- Model maintenance
- Security
- Software development
However, for organizations with large workloads, running a model themselves can potentially reduce dependence on an external AI provider.
Open Weight Does Not Automatically Mean Better
It’s important not to assume that open-weight models are always better than proprietary models.
Closed models can have significant advantages in areas such as:
- Model performance
- Infrastructure
- Reliability
- Safety systems
- Developer tools
- Multimodal capabilities
- Product integration
- Technical support
Open-weight models can offer greater control and flexibility, but they also shift more responsibility to the user.
The organization deploying the model may have to manage infrastructure, security, updates, monitoring, optimization, and compliance itself.
The Geopolitical AI Competition
The debate over open and closed AI is no longer purely technical.
It is becoming part of a broader competition over who controls the global AI ecosystem.
The United States remains home to several of the world’s largest AI companies and leading AI infrastructure providers. At the same time, Chinese companies have demonstrated that openly distributed models can compete for developer attention and global adoption.
Meta’s decision to release open-weight models has also shown that the strategy is not limited to Chinese companies.
This creates two competing approaches.
Closed AI emphasizes centralized infrastructure, proprietary technology, controlled access, and recurring revenue from AI services.
Open-weight AI emphasizes distribution, developer adoption, local deployment, customization, and broader access to model technology.
Neither approach is guaranteed to win.
The future could instead involve both models operating alongside each other, with companies choosing between cloud-based proprietary systems and self-hosted open-weight models depending on their technical, financial, and security requirements.
What Does Open-Weight AI Mean for Businesses?
For businesses, the biggest advantage of an open-weight model is control.
A company can potentially decide where the model runs, how it is configured, what data it processes, and how it integrates with internal systems.
This can be especially useful when an organization needs to customize an AI model for a specialized workflow.
For example, a company could deploy a model internally and fine-tune it for its own documentation, customer-support processes, software development tasks, or industry-specific terminology.
However, businesses also need to consider the total cost of ownership.
Running an AI model internally means taking responsibility for infrastructure, maintenance, security, upgrades, and performance optimization.
For some companies, paying for an API remains simpler and more cost-effective.
The difference between closed AI and open-weight AI comes down to more than whether users can download a model.
Closed AI keeps the model’s core technology under the control of the company that developed it, with users typically accessing the system through an application or API.
Open-weight AI makes trained model parameters available, giving developers and organizations the option to run the model themselves and, where permitted, customize it for their needs.
Open-weight models can offer greater control over data, deployment, and customization, but they also require more technical expertise and infrastructure.
The growing popularity of models from companies such as Alibaba, DeepSeek, Moonshot AI, and Meta shows that open distribution is becoming an important part of the AI industry.
For the United States and China, the issue is also about technological influence. The companies and countries whose models become widely adopted could gain an advantage in shaping the tools, infrastructure, and developer ecosystems that define the next generation of AI.
Ultimately, the future of AI may not be a simple choice between open and closed models. Both approaches have clear advantages, and the winning strategy could depend on the specific needs of developers, businesses, governments, and individual users.
