Why Big Tech is giving away open-weight AI models and what the companies want in return

Why in news?

Big Tech firms released several open-weight AI models in October 2026, reviving debate on motives, risks and relevance for India.

UPSC Relevance 

Prelims – Science and Technology: AI terms like model weights, open-weight vs open-source, APIs, LLMs.

Mains GS3 – Science and Technology – developments and their applications in everyday life; awareness in the fields of IT and computers; indigenisation of technology. 

Understanding the basics

• AI model – A system that learns patterns from huge amounts of data during training and then uses them to answer questions, write code, read images or forecast trends.
• Weights (parameters) – The numerical values learnt during training. They decide how the model turns an input into an output. Think of a chef who has learnt from thousands of recipes – the weights are his acquired sense of what to mix and in what proportion.
• Closed or proprietary model – Weights stay with the company. Users reach the model only through a website, app or an Application Programming Interface (API), and the provider controls access and charges for usage.
• Open-weight model – Trained weights are made downloadable. Anyone can run the model on their own computers or servers, fine-tune it for a task or build it into an application.
• Open-weight is not fully open-source – A firm may release weights without sharing the training code or details of training data. Licences may also limit use. The Open Source Initiative’s Open Source AI Definition (2024) expects code, data information and weights to be shared, which most ‘open’ models do not meet.
• Not a new idea – OpenAI released GPT-2 weights in 2019. What has changed is the capability now available freely, and the rise of Chinese developers like DeepSeek and Alibaba as serious competitors.

Why are companies giving their models away free?

If the model is free, where does the money come from? The answer depends on what each firm actually sells.

• Building an ecosystem – Developers who build apps and tools on a model make it easier to use, which pulls in more developers. This self-reinforcing cycle is the network effect.
• Becoming the industry standard – Meta’s Mark Zuckerberg (essay ‘Open Source AI Is the Path Forward’, July 2024) compared Llama with Linux – he wants Llama to become the default, so Meta is not dependent on rivals’ closed technology.
• Selling the complements – NVIDIA sells chips and software, so wider AI use raises demand for computing. Alibaba’s Qwen sits next to its cloud services. Mistral offers paid hosting, enterprise support and private deployments.
• Brand, talent and research – In a 2024 submission to the US NTIA, OpenAI listed brand recognition, recruitment and attracting developers as reasons to release weights.
• Shaping how AI is used – A widely used model sets the norms, tools and formats that the rest of the industry follows – a long-term influence that cannot be bought later.

The geopolitical angle

• China’s challenge – DeepSeek-R1 (January 2025, MIT licence) and Alibaba’s Qwen family showed that strong models could be built cheaply and shared openly. Developers pick on cost and performance, so a cheaper Chinese model can replace an American or European one.
• ‘Democratic AI’ vs ‘autocratic AI’ – When OpenAI released its open-weight models in August 2025, it argued they would help partner countries build on ‘US-led rails’. AI spread is now openly treated as a strategic contest, much like 5G or semiconductors.
• Sovereign AI – Countries increasingly want AI that runs on their own infrastructure, understands their languages and keeps their data at home. Open weights make this easier but the base technology may still come from abroad.
• Chips remain the choke point – Weights may be free, but running models needs advanced GPUs, a market dominated by a few firms and shaped by US export controls.

Benefits of open-weight models

• Lower entry barrier – Startups, universities and governments in developing countries can use advanced AI without building a model from scratch, which costs huge sums.
• Data privacy and security – Sensitive data in health, defence or governance can be processed on in-house servers instead of being sent to a foreign company’s servers.
• Local adaptation – Models can be fine-tuned for Indian languages, dialects, legal texts and local needs.
• No vendor lock-in – Reduces dependence on one provider whose prices, terms or availability may change.
• Transparency and research – Researchers can study, test and audit model behaviour, which helps safety work and builds trust.
• Competition – Keeps prices of closed services in check and spreads innovation.

Concerns and challenges

• Misuse is hard to control – Users can strip out safety guardrails and use models for deepfakes, fraud, cyberattacks or disinformation. Anthropic CEO Dario Amodei has said open models can be a public good but cautioned against releasing ones with dangerous capabilities.
• Irreversible release – Once weights spread, they cannot be recalled. With API access, a provider can monitor and restrict use; with downloads it cannot.
• Accountability gap – When a modified model causes harm, it is unclear whether the original developer, the modifier or the deployer is responsible.
• ‘Openwashing’ – Some licences carry restrictions – for instance, Llama’s licence requires a separate permission for services with over 700 million monthly users. ‘Open’ does not always mean free to use in every way.
• Hidden costs – Hardware, maintenance, security and skilled staff still cost money. Running a model in-house is not always cheaper than a hosted service.
• Bias and cultural mismatch – Models trained mostly on English and Western data may perform poorly or carry bias in Indian contexts.
• Dependence in another form – Building only on foreign base models can create a softer kind of reliance on technology choices made elsewhere.

Where does India stand?

• Different goals – Commercial firms look for revenue from apps and enterprise services, while public research projects focus on language access and domestic capability.

Global regulatory approaches

• European Union – The AI Act (2024) gives open-source general-purpose models some relief from documentation duties, but not if they pose ‘systemic risk’.
• United States – Largely market-led; open-weight release is encouraged as a tool to spread American technology.
• China – Firms are pushing openly released models, while content and security rules apply to public-facing AI services.

Way forward

• Build sovereign capacity, not just adopt – Use open weights as a starting point while investing in Indian data (AIKosh), compute and home-grown models so that dependence does not simply shift form.
• Risk-based openness – Release weights in stages and test frontier models for dangerous capabilities before release; the IndiaAI Safety Institute can lead such evaluations.
• Clear liability rules – Spell out responsibilities of developers, modifiers and deployers, especially for deepfakes and fraud.
• Compute as public good – Expand subsidised GPU access for startups, researchers and public institutions so that open models are actually usable.
• DPI model for AI – Like UPI and Aadhaar, offer open Indian models as shared building blocks for public services in health, agriculture and education.
• Global South partnerships – Share Indian-language and low-cost AI solutions with other developing countries, strengthening India’s voice in global AI rule-making.
• Skills – Train engineers to fine-tune, deploy and secure open models, not just use ready-made chatbots.

Open weights widen access and competition, but they are as much a business and strategic tool as a gift. For India, the real test is whether these models build lasting domestic capability.

Practice MCQ

Q1. Consider the following statements regarding open-weight Artificial Intelligence (AI) models:

1. Release of the model weights necessarily includes release of the complete training data and source code.

2. Once the weights are widely distributed, the original developer cannot withdraw them.

3. Such models can be run on a user’s own infrastructure without accessing the developer’s API.

How many of the above statements are correct?

(a) Only one        (b) Only two        (c) All three        (d) None

Answer: (b). Statement 1 is incorrect – firms may release weights without the training code or data details, which is why open-weight is not the same as fully open-source.

Q2. Consider the following statements:

Statement-I: Several technology companies are releasing their AI models as open-weight models free of cost.

Statement-II: Wider adoption of a model can create a network effect that increases demand for the company’s cloud, hardware and enterprise services.

Which one of the following is correct in respect of the above statements?

(a) Both Statement-I and Statement-II are correct and Statement-II explains Statement-I

(b) Both Statement-I and Statement-II are correct but Statement-II does not explain Statement-I

(c) Statement-I is correct but Statement-II is incorrect

(d) Statement-I is incorrect but Statement-II is correct

Answer: (a). Firms give away models to build ecosystems around them and earn from complementary services like hosting, chips and enterprise support.

Mains Practice Question

Q. “Open-weight AI models are projected as tools for democratising technology, but they also serve corporate and geopolitical interests.” Examine. What opportunities and risks do they present for India’s pursuit of sovereign AI? (15 marks, 250 words)

PDF