
@SarvamAI’s Token Factory announcement has drawn criticism with some confounding Sovereign AI with Frontier AI. The @OfficialINDIAai Mission allocation was directed at indigenous model development. Sarvam drew down on that allocation to train and open-source Sarvam-30B and Sarvam-105B from scratch in India under Apache 2.0. These are models developed and trained in India. They are sovereign, but not frontier. While the Mission allocation supported training of models at the 30B–105B scale, reaching the current frontier requires substantially larger compute, data, and capital than that allocation held. Sarvam-105B is capable for its size and on Indian language tasks, albeit it does not compete at the leading edge of general reasoning, coding or scale. Token Factory’s decision to serve stronger external open-weight models follows on from that gap, not from complete absence of indigenisation. Criticism concerning the Chinese models: Token Factory makes Sarvam’s 105B available alongside Chinese open-weight models as DeepSeek, GLM, and Kimi that lead the open-weight frontier in reasoning and coding. Zero data retention, runs locally on Indian GPUs with no telemetry or data sent back to the original developers and with zero data retention available on request. While hosting 3rd party open-weights is standard industry practice, the material difference here is domestic data residency enabling Indian enterprises local regulatory and legal compliance without relying on overseas APIs. Overall, sovereign models were built and released, and they are not frontier. The substantive question will be to balance massive capital needs and further original model development against local inference of stronger open-weight models.
Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability. For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.













