Act now to secure India’s access to the latest AI models

Most AI developers prefer to use open-weight models to build their solutions because they can be fine-tuned to suit specific requirements.

- RAHUL MATTHAN

Take, for example, DeepSeek, the open-weight Chinese AI model. Perplexity was able to fine-tune it to remove all China-specific bias, as a result of which its users were able to get honest answers to questions about the 1989 protests at Tiananmen Square. This is not possible when the same questions are asked of the DeepSeek API (or application programming interface).

Perhaps more importantly, open-weight models—these are not ‘open-source’ astraditionally understood but allow some amount of adjustment—can be downloaded and deployed in their own compute environment, ensuring that developers are not dependent on big AI labs for continued access to the models they need. This gives companies developing retail applications a level of freedom they otherwise lack.

I used to believe that open weight models were the way to go until I read the documentation accompanying GPT-OSS, OpenAl's first open-weight model since GPT-2. In the safety paper it released, OpenAI stated that it had intentionally kept the model's capabilities below the current frontier, as that was the only way to ensure that bad actors would not be able to use it for nefarious purposes. While I am all for ensuring that AI models are safe, Iworry that if there is no other way to guarantee model safety, the long-term implications for Indian developers will be grim.