Beam. Muse Spark Open. Nemotron 3.5 Super VL. Nemotron 3.5 Bolt 74B-A9B. Timestral.
These may sound like characters from a science-fiction franchise. They are, in fact, names of open-weight AI models announced or released by technology companies in the United States and Europe in October.
Reflection AI released Beam’s weights on October 16. Meta released Muse Spark Open, while NVIDIA made two Nemotron models available, on October 15. France-based Mistral AI also announced Timestral, a family of models designed for time-series forecasting.
Companies are making models available for different tasks, from writing software and operating AI agents to interpreting images and forecasting time-series data. But why spend heavily to develop AI models and then allow others to download and adapt them? If the models are free to download, where does the money come from?
The answer lies in a race to attract developers, shape how AI is used and earn money from the businesses, services and computing infrastructure built around these models.
What is an open-weight model?
AI models learn patterns from data during training. Their trained parameters, called weights, are numerical values that help determine how the models process inputs and generate responses.
Think of an AI model as a chef who has learnt from thousands of recipes. During training, the model learns patterns that help it decide what ingredients to combine and in what proportions to produce a particular dish. Its weights are the numerical values that shape these decisions. Making those weights available is like sharing the chef’s learned knowledge with other chefs, who can then adapt it to their own needs and create dishes of their own.
Most people access proprietary AI models through websites or apps, while developers may access them through application programming interfaces (APIs). The provider controls access and often charges for usage.
An open-weight model makes its trained parameters available for download. Developers can run it on compatible computers or servers, adapt it for particular tasks, or integrate it into their applications.
This can give businesses and individual users greater control over how they deploy and customise AI models, and reduce their dependence on a single provider. However, open-weight does not automatically mean open-source in the fullest sense. A company may release the weights without publishing its training code or sufficient information about the data used to train the model. Licences may also restrict certain uses.
Open-weight AI is not new. OpenAI released GPT-2’s weights in 2019. What has changed is the level of capability available through downloadable models and the emergence of Chinese developers such as DeepSeek as serious competitors, prompting a broader contest over who will shape the AI ecosystem.
The Chinese competition
The growing availability of Chinese models has intensified competition in the open-weight AI market. Companies including DeepSeek and Alibaba have released models whose weights developers can download and adapt. DeepSeek-R1, released in January 2025 under the MIT licence, made a reasoning model available for others to use, modify and commercialise. Alibaba’s Qwen family includes models of different sizes for a range of applications.
Developers choose models based on performance, cost, hardware requirements and licensing. If a Chinese model meets their needs at a lower overall cost, they may choose it over a proprietary American or European service. As more developers build applications and tools around a model, its ecosystem can make it more attractive to others.
OpenAI made the geopolitical argument explicit when it released its open-weight models in August 2025. It said the models could help partner countries build AI infrastructure on what it called “US-led rails”, giving them an incentive to choose “democratic AI” over “autocratic AI”. This is the company’s stated position, but it illustrates how the spread of AI technology is also viewed as a strategic contest.
If the model is free, where does the money come from?
The answer depends on what the company sells. Meta chief executive Mark Zuckerberg outlined his approach in a July 2024 essay, Open Source AI Is the Path Forward. Drawing a comparison with Linux, he argued that releasing Llama’s weights could encourage companies to develop tools, integrations and hardware optimisations around the models, helping build a broader ecosystem.
Mr. Zuckerberg said selling access to AI models was not Meta’s business model, which centred on building products and services. He argued that releasing Llama would help Meta avoid dependence on competitors’ proprietary technology while encouraging the development of a broader ecosystem of tools and services around its models.
For NVIDIA, the commercial interest is different. It sells computing hardware and related software. Wider adoption of AI applications could increase demand for computing power, potentially benefiting NVIDIA, although developers can run open models on competing hardware. Revenue gains from individual model releases are not guaranteed.
Model developers such as Mistral can offer paid hosting, enterprise support, customised models and private deployments. Alibaba’s Qwen models sit alongside its cloud and AI services.
A developer may download a model’s weights at no cost but pay a provider to host and run it reliably at scale. Alternatively, the developer may run it independently and never pay its creator. The business bet is that wider adoption will generate enough revenue or strategic value elsewhere to justify the cost of developing and releasing the model.
The fight to shape the AI ecosystem
The longer-term prize is influence. A widely used model can attract developers who build applications and tools around it. These additions can make the model easier to deploy and use, potentially attracting more developers. This reinforcing cycle is known as a network effect.
Distribution matters too. Making model weights downloadable is only one way to encourage adoption. Offering access through APIs, model hubs and inference platforms can make models easier for developers to discover, test and integrate into applications.
Mr. Zuckerberg has argued that Meta wants Llama to become an industry standard. In a 2024 submission to the U.S. National Telecommunications and Information Administration, OpenAI identified brand recognition, recruitment and attracting developers as reasons companies release model weights. It also warned that widely distributing weights can make misuse harder to control, since users can modify models and remove safeguards.
The strategy is therefore not simply about giving technology away. It is about encouraging developers to build on a company’s models, potentially increasing demand for its infrastructure, tools and services as AI applications evolve.
The risks of opening up models
Not everyone believes more openness is always better. Anthropic chief executive Dario Amodei has argued that open-weight models can be a public good, but warned that releasing models with dangerous capabilities can make harmful uses harder to prevent. Once weights are widely distributed, they cannot simply be withdrawn.
OpenAI has made a similar argument saying that access through an API allows a provider to monitor use and restrict access, while downloaded weights can be run independently, beyond the original developer’s control.
The trade-off is that open weights can widen access and encourage competition, but reduce the original developer’s ability to intervene when a model is misused.
What it means for India
For Indian businesses, open-weight models offer an alternative to relying exclusively on foreign AI services. Organisations can adapt models to Indian languages and local requirements, and may choose to run them on their own infrastructure, depending on their computing resources and the model’s licence.
India has its own efforts too. Sarvam AI has released models including Sarvam 30B and Sarvam 105B, while AI4Bharat at IIT Madras has made Indian-language models and tools available to developers.
BharatGen has also released models for Indian languages and domain-specific applications, including agriculture, law and Ayurveda. The objectives, however, vary. Commercial companies may seek revenue through applications and enterprise services, while research initiatives may prioritise language access and domestic technical capability.
Open weights alone do not guarantee lower costs. Hardware, maintenance, performance and licensing still matter, and organisations must assess whether running a model themselves is more economical than using a hosted service.