Sarvam serves DeepSeek V4.1 Flash through its Token Factory inference platform and says it handles m...
Sarvam's Token Factory serves DeepSeek V4.1 Flash and other open-weight models on Indian GPUs, raising questions about how its sovereign-AI pitch fits a hosting business.
Sarvam serves DeepSeek V4.1 Flash through its Token Factory inference platform and says it handles more than 450 billion tokens a day on that model. Sarvam also says 95.7% of prompt tokens on it are served from cache. Token Factory runs open-weight models on Sarvam's own GPU clusters in India behind one OpenAI-compatible API. Other models named at launch include Sarvam 105B, Gemma 4 31B, GLM 5.3 and Kimi K3.1 Pro. Sarvam 105B costs Rs 15 per million input tokens, Rs 60 per million output tokens and Rs 5 per million cached input tokens. Sarvam says it matches or beats leading global providers on speed and cost for most hosted models. Those figures are the company's own.
The product puts a sovereign-AI company in the business of hosting other developers' models. Here, sovereignty means in-country compute and data residency. Sarvam says requests do not go to the model developers' overseas APIs. The market is crowded: Krutrim has hosted DeepSeek on Indian servers since January 2025, and OpenAI and Anthropic now offer in-country inference for some models. Our earlier coverage on 2026-11-18 noted that Token Factory pairs local inference with mostly foreign-developed models.
For builders, Indian teams with data-residency rules, such as banks, can switch models by changing one field in the API call. They can also choose dedicated GPU capacity or an on-premise deployment of the same software. For investors, the open question is how much traffic goes to third-party models rather than Sarvam 105B. That split will show whether Token Factory builds Sarvam's own model business or mainly resells hosting.



