
In this article
Why in the news
Indian AI company Sarvam launched Token Factory on 16 November 2026. The company describes it as an inference platform serving its own and other organisations’ open-weight models from GPU clusters in India. The exam issue goes beyond a particular chatbot: developing a model, running a foreign model locally and controlling user data are different capabilities. [1]
Training, inference and APIs
Training learns a model’s parameters from examples. Inference uses a trained model to produce an output for a new request. GPUs are computing chips useful for the parallel mathematical work involved in AI. Sarvam describes Token Factory as an inference stack, rather than one new model.
An API is a defined way for software systems to exchange requests and responses. Sarvam says its service provides several models through an OpenAI-compatible API. Compatible describes the interface here; it does not establish ownership or endorsement by another company. [1]
A model’s origin is not its hosting location
The launch lists Sarvam 105B alongside models from the GLM, DeepSeek, Kimi and Gemma families. Sarvam says it serves them from GPU clusters in India. This illustrates the difference between where a model originated and where it runs. Hosting a foreign-origin model in India does not make it Indian-trained. [1]
Open-weight means learned model parameters are available. It does not by itself mean every training example is public or all uses are unrestricted. The relevant licence still governs permitted use.
Four tests of capability
Model development: Can domestic institutions develop and improve models for Indian languages and needs?
Computing capacity: Are chips, electricity, networks and operating skills reliably available?
Data governance: Where are requests, logs and outputs stored? Who can access them, and for how long are they retained? Local hosting alone does not answer all these questions.
Freedom to switch: Access to several models may reduce reliance on one provider, but API dependencies, licences and service terms still matter. These are criteria for analysis, not a completed security assessment of this platform.
Opportunity and caution
A local inference service may help a small business build an AI application without installing its own large computing system. Developing domestic multilingual AI capability is also a stated policy objective. A February 2026 PIB backgrounder places Sarvam within India’s indigenous, multilingual AI ecosystem. That is not proof of government security certification or government financing for a particular November commercial service. [2]
The launch’s cost and speed figures are company-reported tests. Results depend on the model, request length, concurrent users and testing method. This article therefore does not present them as an independently verified performance ranking. [1]
Way forward
Policy can assess domestic research and the use of global open-weight models separately. Sensitive deployments need checks of data flows, independent security tests, licences, procurement terms and the ability to change models. The question is not only where a model was made, but who has meaningful control over the deployed system and how that control is held accountable.
Prelims Practice Questions
Question 1
Consider the following statements about AI platforms:
Statement 1: Inference uses a trained model to produce an output for a new input.
Statement 2: Every model hosted in India was necessarily trained in India.
Statement 3: An API can define how software systems exchange requests and responses.
Which of the statements given above are correct?
Answer and Explanation
Correct answer: (a)
Statement 1 is correct: training and inference are different stages.
Statement 2 is incorrect: the hosting location does not establish where a model was trained or originated.
Statement 3 is correct: an API defines an interface for software communication.
Question 2
Which best describes Token Factory in its launch announcement?
Answer and Explanation
Correct answer: (b)
The company describes an inference stack for its own and other open-weight models. Factory in its name does not make it a chip-manufacturing plant.
Question 3
Which is the soundest conclusion about an open-weight model?
Answer and Explanation
Correct answer: (c)
Access to model weights differs from access to all training data or unrestricted usage rights. Licences and service terms must be checked separately.
Mains Practice Question
“AI sovereignty cannot be measured only by a model’s country of origin.” Examine with reference to local inference platforms. (150 words, 10 marks)
Answer Outline
AI sovereignty involves sufficient capability and control over research, computing and data decisions. A model’s origin is one dimension, not the complete test.
Domestic models can build capacity for Indian languages and specific needs. Local inference infrastructure may improve service availability, operating skills and business access. Yet running a foreign model on domestic servers does not automatically create control over training data, licensing or key technology.
Assessment should cover data flows, log retention, security tests, licences, computing availability and the ability to switch providers. Company-reported speed or cost claims cannot replace independent tests. Public procurement should set conditions according to data sensitivity and the accountability required.
Domestic innovation can coexist with careful use of global technology. Meaningful capacity, safe deployment and verifiable accountability are better tests than a label alone.
Frequently Asked Questions (FAQ)
Does OpenAI-compatible mean the service is an OpenAI model?
No. Here it is a claim about API-interface compatibility, not proof of model ownership or endorsement by another company.
Does local hosting guarantee complete data security?
No. Data collection, logs, access rights, service terms and security testing also require scrutiny.
Are the launch’s performance figures an independent ranking?
No. They are company-reported tests. Results from different conditions cannot be treated as equivalent without a common testing method.
Sources
- https://www.sarvam.ai/blogs/sarvam-token-factory
- https://www.pib.gov.in/PressReleasePage.aspx?PRID=2231169&lang=1®=3
Download this article: PDF
