Op-ed: China's cheap open-weight AI models may not stay open for long
Indian startups build on Qwen and DeepSeek because they are cheap. What happens if the tap is turned down?
Published 16 September 2026. Written by Pratidin from the reports linked at the end; every fact checked by a separate review before publishing. How we work
An opinion article in The Hindu on 16 September 2026 argues that China's lead in cheap, open-weight artificial intelligence may not last and that India should prepare. It says Indian startups are increasingly adopting Chinese open-weight large language models (LLMs) such as Qwen, DeepSeek and Kimi, whose trained parameters, or weights, can be downloaded and run locally. According to the article, these models perform nearly as well as American frontier models, with a lag of about six months, and China had 820 LLMs registered by early 2026. It attributes low costs, such as a reported $294,000 for DeepSeek's R1, to distillation from American models, a technique in which a smaller model learns from a larger model's answers.
The writer reads China's open release strategy as a way to expand global influence, commoditise AI and build demand for complementary services, and projects that open access will face graduated restrictions around late 2028. The recommendation is not full self-sufficiency. Instead, India should build model-agnostic systems that make it cheap to switch models, focus on applications, industrial data and domain-specific fine-tuning, and help shape norms for open-weight models in multilateral forums. The context is DeepSeek's R1, released on 20 January 2025 under the MIT licence, whose rise wiped about $600 billion off Nvidia's market value, and India's IndiaAI Mission, approved in 2024 with over ₹10,300 crore over five years.
The article's timelines are forecasts and should be quoted as the author's view. There are counterpoints. Weights already released under permissive licences can be kept and run, so any restriction would mainly affect future models and services. Cost comparisons are disputed: DeepSeek's reported $5.6 million figure for its V3 model was criticised as covering only part of the true cost. And dependence on any single foreign provider, Chinese or American, raises similar risks of changed terms, embedded bias and data security, which strengthens the case for the flexible, application-led approach the article proposes and for the shared compute being built under IndiaAI, whose first round in 2025 approved 10 providers to supply 18,693 GPUs.
Prelims facts
- An open-weight AI model releases its trained weights for download; its training data and code may still be withheld.
- The op-ed says Chinese models such as Qwen, DeepSeek and Kimi trail U.S. frontier models by about six months and projects restrictions around late 2028.
- Knowledge distillation trains a smaller model on a larger model's outputs, cutting cost.
- DeepSeek-R1 was released on 20 January 2025 under the MIT licence.
- The IndiaAI Mission, approved in 2024, has over ₹10,300 crore over five years and its first compute round in 2025 approved 10 providers to supply 18,693 GPUs.
Quick recall
- Which Chinese open-weight models does the op-ed say Indian startups are adopting?
- Qwen, DeepSeek and Kimi.
- What is stored in an AI model's 'weights'?
- The numbers (parameters) that hold what the model learned in training.
- What is knowledge distillation?
- Training a smaller student model to imitate a larger teacher model's outputs.
- When was DeepSeek-R1 released, and under which licence?
- 20 January 2025, under the MIT licence.
- Who founded DeepSeek?
- Liang Wenfeng, in July 2023; it is backed by the hedge fund High-Flyer.
- Outlay of the IndiaAI Mission?
- Over ₹10,300 crore over five years; approved in 2024.
- How many GPUs did the first round of the IndiaAI common computing facility approve?
- 18,693, from 10 providers (2025).
- When does the op-ed expect restrictions on Chinese open-weight access?
- Graduated restrictions around late 2028 (the author's projection).
Prelims practice question
In the context of artificial intelligence, the term 'open-weight model' refers to a model whose:
- training data must be made public by law
- trained parameters are released for anyone to download and run
- outputs are free of copyright
- source code is certified by an international standards body
Show answer
Answer: (b) trained parameters are released for anyone to download and run. An open-weight model publishes its trained parameters (weights), so it can be downloaded, run locally and fine-tuned. Releasing the weights does not by itself mean the training data or full code is public, and it has nothing to do with copyright status or certification.
Use this in UPSC Mains: previous-year questions
Recurring theme: Artificial intelligence, technology sovereignty and India-China competition
- How to use this
Use it to define AI models, weights and distillation simply; the dependence debate supports the data security part of the privacy discussion.
- Large language models learn trained parameters called weights; open-weight models such as Qwen, DeepSeek and Kimi release these weights for download, though training data and code may be withheld.
- Knowledge distillation trains a smaller model on a larger model's answers, cutting cost; an op-ed attributes DeepSeek R1's reported $294,000 cost to it.
- Dependence on any foreign model raises risks of embedded bias, limited auditability without training data, and data security in sensitive sectors, arguing for security reviews.
- How to use this
Indian AI startups' reliance on Chinese open-weight models illustrates a global challenge for startups, and IndiaAI's shared compute is a domestic remedy.
- An op-ed (The Hindu, 16 September 2026) says Indian startups increasingly build on Chinese models such as Qwen and DeepSeek and projects that access may face restrictions around late 2028.
- The IndiaAI Mission, approved in 2024 with over ₹10,300 crore over five years, approved 10 providers to supply 18,693 GPUs in its first compute round in 2025.
- The op-ed recommends model-agnostic systems that make switching cheap, and a focus on applications, industrial data and domain-specific fine-tuning.
Reducing reliance on Chinese AI models is a new layer of the supply-chain de-risking the question describes.
Mains practice question
Cheap open-weight AI models from abroad offer Indian startups a head start but also create strategic dependence. Discuss how India should balance the two. (150 words)
Model answer
An op-ed in The Hindu (16 September 2026) notes that Indian startups increasingly build on Chinese open-weight models such as Qwen and DeepSeek, and warns that access may narrow.
Benefits
- Low cost and no per-use fees; weights can run on local servers.
- Fine-tuning on Indian data and languages.
- Faster product launches for startups.
Risks
- Terms or future versions controlled abroad; the author projects restrictions around late 2028.
- Embedded bias and limited auditability without training data.
- Data security in sensitive sectors.
Balancing approach
- Model-agnostic design so products can switch models cheaply.
- Build where value lies: applications, industrial data, domain fine-tuning.
- IndiaAI Mission: shared compute (18,693 GPUs approved in its first round in 2025) and home-grown models such as BharatGen.
- Norm-setting on open-weight AI in multilateral forums.
- Security reviews for any foreign model in critical uses.
Flexibility, not isolation, lets India use today's cheap models while building its own capacity.
The basics
Why this matters
Many Indian startups now build products on top of AI models made by someone else. An opinion article in The Hindu on 16 September 2026 argues that Chinese open-weight models such as Qwen, DeepSeek and Kimi have become the cheap default, and that India should plan for the day that access narrows.
Open, open-weight and closed
A large language model (LLM) is a program trained on huge amounts of text to predict and generate language. What it learns is stored in its "weights", billions of numbers. Who can see and use those numbers decides how "open" a model is. See Open-weight versus open-source AI.
- Weights stay with the company
- Users pay per use and send data to the provider
- Access can be changed or withdrawn at any time
- Weights can be downloaded and run locally
- Can be fine-tuned on one's own data
- Training data and full recipe may still be secret
How cheap models get made
Training a frontier model from scratch is costly. One shortcut is Knowledge distillation, in which a smaller model learns from a larger model's answers. The article says low costs such as a reported $294,000 for DeepSeek's R1 are due to distillation from American models.
- 1Pre-trainingThe model learns language patterns from a vast text corpus
- 2Distillation or fine-tuningIt is refined, sometimes by learning from a stronger teacher model
- 3ReleaseWeights are published (open-weight) or kept behind an API (closed)
- 4AdaptationA startup fine-tunes the model on its own domain data
- 5ApplicationThe tuned model powers a chatbot, search tool or industrial system
The article's warning
The writer says Chinese LLMs perform nearly as well as American frontier models, with a lag of about six months, and that China had 820 LLMs registered by early 2026. The motive, in the article's reading, is influence and market share, and it projects graduated restrictions around late 2028. These are forecasts, not facts, and should be quoted as the author's view. The January 2025 release of DeepSeek and the January 2025 shock showed how quickly a cheap Chinese model could move markets.
India's options
The article advises against chasing complete self-sufficiency. It suggests model-agnostic architectures that make it cheap to switch models, a focus on applications, industrial data and domain-specific fine-tuning, and a role in shaping open-weight norms in multilateral forums. India's own effort runs through the IndiaAI Mission.
Go deeper
In one line: An op-ed argues that Indian startups relying on cheap Chinese open-weight AI models should build flexibility now, because that openness may be restricted later.
Why it matters for UPSC
It links GS3 science and technology (AI, indigenisation of technology) with GS2 (China, technology geopolitics, global governance of AI).
The core idea
Most AI products are built on foundation models. Chinese firms release many of theirs as open-weight models; the difference is explained in Open-weight versus open-source AI. Some of their low cost comes from Knowledge distillation, according to the article. The market shock caused by DeepSeek and the January 2025 shock showed their reach. India's own response is the IndiaAI Mission, which funds compute and home-grown models.
Numbers and dates to remember
- Models named in the article: Qwen, DeepSeek, Kimi
- The article: about six months' lag behind U.S. frontier models; 820 LLMs registered in China by early 2026; restrictions projected around late 2028
- DeepSeek R1 released 20 January 2025 under the MIT licence
- DeepSeek founded in July 2023 by Liang Wenfeng, backed by the hedge fund High-Flyer
- IndiaAI Mission approved in 2024: over ₹10,300 crore over five years; 18,693 GPUs approved in the first compute round (2025)
Where to go next
- Open-weight versus open-source AI: what "open" really means for a model
- Knowledge distillation: how smaller models learn from bigger ones
- DeepSeek and the January 2025 shock: the release that changed the AI cost debate
- IndiaAI Mission: India's own compute and model programme
Go deeper: dependence, sovereignty and the smart middle path
The article's case. Cheap open-weight models are a gift today, but gifts from a strategic rival come with terms. The writer sees China's release strategy as a way to expand influence, commoditise AI and build demand for complementary services, and projects graduated restrictions around late 2028. If Indian products are hard-wired to one foreign model, a policy change abroad could raise costs or cut access. So the article recommends model-agnostic design, investment in applications and industrial data, and domain-specific fine-tuning, rather than the costly goal of complete self-sufficiency.
Points that complicate the warning. Weights released under a permissive licence, such as the MIT licence used for DeepSeek's models since January 2025, can be kept and run locally; a future restriction would bite on new versions, updates and services, not on files already downloaded. The article's six-month-lag claim and its 2028 timeline are the author's estimates. And the cost story is contested: DeepSeek's reported $5.6 million figure for its V3 model was criticised as covering only part of the true cost, and the article itself credits Knowledge distillation from American models for low costs.
Other risks of dependence. Open-weight models can carry the values and blind spots of their training data. For sensitive uses, security and data-protection reviews matter whichever country a model comes from. The distinction in Open-weight versus open-source AI also matters: without training data, users cannot fully audit a model.
Where India stands. The IndiaAI Mission funds shared compute and home-grown foundation models such as BharatGen. The lesson of DeepSeek and the January 2025 shock is that efficiency, not only scale, can close gaps. A balanced path is to keep options open across U.S., Chinese and Indian models while building Indian data, talent and applications where the long-term value lies.
Open-weight versus open-source AI
what 'open' really means for a model
In one line: An open-weight model publishes its trained parameters for anyone to download; a fully open-source model would also share what is needed to rebuild it.
The spectrum of openness
- Closed models: reached only through an API; the company keeps the weights and can change prices or terms.
- Open-weight models: the weights are released, so anyone can run the model on their own machines and fine-tune it. The training data and full training code may still be withheld.
- Fully open-source models: weights, code and enough information about data to reproduce and audit the model.
Why it matters for India
Open weights let Indian startups avoid per-use fees and keep sensitive data on their own servers. But without the training data, users cannot fully check a model's biases, and future versions depend on the publisher's choices, the risk the op-ed highlights.
Where to go next
Knowledge distillation
how smaller models learn from bigger ones
In one line: Knowledge distillation trains a smaller "student" model to imitate the answers of a larger "teacher" model.
How it works
Instead of learning only from raw text, the student is shown many questions along with the teacher's responses and learns to reproduce them. Because the teacher has already done the hard learning, the student can reach good performance with far less computing power.
Why it is in the news
The op-ed says the low training costs of some Chinese models, such as a reported $294,000 for DeepSeek's R1, are due to distillation from American models. This raises two debates: whether cost comparisons are fair when one model builds on another's work, and whether using a rival's outputs this way breaches terms of service. For developing countries, distillation is also an opportunity, since it lowers the cost of capable models.
Where to go next
DeepSeek and the January 2025 shock
the release that changed the AI cost debate
In one line: DeepSeek, a Chinese AI firm, released its R1 model in January 2025, and its low cost and open weights shook global markets.
The company
DeepSeek was founded in July 2023 by Liang Wenfeng and is backed by the Chinese hedge fund High-Flyer. Since January 2025 it has released models under the MIT licence, a permissive licence that allows reuse.
The shock
DeepSeek-R1 was released on 20 January 2025. By 27 January 2025 its app had overtaken ChatGPT as the most downloaded free app on the U.S. iOS App Store, and Nvidia lost about $600 billion in market value, the largest single-company fall in U.S. history. DeepSeek's reported $5.6 million training cost for its V3 model was questioned as covering only part of the true cost. The episode showed that efficiency could narrow the gap with frontier labs.
Where to go next
IndiaAI Mission
India's own compute and model programme
In one line: The IndiaAI Mission, approved in 2024, is India's programme to build AI compute, datasets and home-grown models.
What it funds
The mission was allocated over ₹10,300 crore over five years. Its centrepiece is a common computing facility; its first round, in 2025, approved 10 companies to supply 18,693 graphics processing units (GPUs), so that startups and researchers can rent compute cheaply. Three AI Centres of Excellence were set up in healthcare, agriculture and sustainable cities, and a Budget announced a new Centre of Excellence for AI in education with ₹500 crore.
Home-grown models
The mission supports Indian foundation models, including large language models suited to Indian needs. BharatGen has been described as the world's first government-funded multimodal LLM initiative. The government has also spoken of developing an Indian GPU within three to five years.
Where to go next
Take the 16 September 2026 quiz: 30 Prelims-style questions with answers