Bengaluru-based AI startup Sarvam has opened Token Factory, a commercial inference service that runs leading open-weight AI models on GPU clusters located in India, offering developers a single OpenAI-compatible API with pricing in rupees.

Source: AlphaSignal

The launch catalogue includes Sarvam's own 105-billion-parameter model, Sarvam 105B, alongside DeepSeek V4.1 Flash 1017 and DeepSeek V4 Pro, GLM 5.3 and GLM 5.3 Flash, Kimi K3 and K3.1 Pro, and Google's Gemma 4 31B.

Introductory pricing puts DeepSeek V4.1 Flash at ₹14 per million input tokens and ₹55 per million output tokens; Sarvam 105B is priced at ₹10 input and ₹25 output per million tokens, with the introductory rates running to the end of 2026.

Sarvam says the service already handles more than 450 billion DeepSeek tokens a day, with 95.7 per cent of prompt tokens served from cache. Those figures are the company's own claims.

Source: BusinessBearings

Why critics are asking questions

The debate centres on the mix: six of the eight launch models come from Chinese developers — DeepSeek, Moonshot AI's Kimi and Z.AI's GLM — alongside Google's Gemma and Sarvam's own model. Critics have asked whether that fits the sovereign-AI mission Sarvam was picked for: in April 2025 the government selected Sarvam under the IndiaAI Mission to build India's own foundational model, and the company received a ₹98.68 crore GPU subsidy for 4,096 H100 graphics processors.

Source: Medianama

Sarvam has pushed back, saying Token Factory runs entirely on privately procured GPUs and does not rely on government subsidies. Co-founder Pratyush Kumar told The Hindu BusinessLine that Sarvam's subsidised GPU access under the IndiaAI Mission ended in March 2026, and that the subsidy supported training open-weight models published on Hugging Face and AIKosh.

The row comes days after Sarvam signed a deal with the Headquarters Integrated Defence Staff to develop and deploy AI applications for the Army, Navy and Air Force — a deal Awaaz Punjab reported on October 27.

For Indian enterprises, the pitch is data residency: prompts run on Sarvam-operated GPUs in India rather than being forwarded to a foreign model provider's overseas API, with zero data retention available on request.