Image: Sarvam AI
Sarvam AI has opened Token Factory to everyone. It is the Bengaluru company's paid service for running AI models on its own GPUs in India, and it serves Sarvam's own models alongside popular open-weight ones from other labs. Sarvam says it matches the leading inference providers on performance and reliability, and pitches the service to enterprises as a way to pick from more models without getting locked in.
Today, we are launching Sarvam Token Factory for everyone.
— Sarvam (@SarvamAI) November 16, 2026
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. pic.twitter.com/RMqEQIFqdC
The catalogue at launch includes Sarvam 105B, GLM 5.3 and GLM 5.3 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro, Kimi K3 and Kimi K3.1 Pro, and Gemma 4 31B, all through one OpenAI-compatible API. Sarvam's product page also lists Qwen 3.8 27B, an embedding model and its own Prism decision model. Prices are in rupees and exclude GST. Sarvam 105B costs ₹10 per million input tokens and ₹25 per million output tokens at an introductory rate, while Kimi K3.1 Pro sits at the top at ₹200 and ₹950.
The launch blog leans hard on speed numbers. Sarvam says Sarvam 105B returns its first token in 0.2 seconds and streams 330 tokens per second for a single user. It says Gemma 4 31B answers a 10,000-token prompt in 0.40 seconds, five times faster than the only other provider serving that model, which it does not name.

Image: Sarvam AI
DeepSeek V4.1 Flash comes in two versions. The Interactive endpoint runs at about 200 tokens per second per user for coding agents and costs ₹25 in and ₹105 out per million tokens. The Standard endpoint is meant for batch work at ₹14 and ₹55. Sarvam's own chart shows Together serving the model faster, at 278 tokens per second, but at a higher price, and shows OpenAI's GPT-6 Luna as cheaper than either Sarvam option.

Image: Sarvam AI
The comparisons need some context. Sarvam ran every test itself, using what it calls a recreation of Artificial Analysis's method, from a client in AWS Mumbai. A service hosted in India has a natural network advantage when measured from India, so the results say less about users elsewhere. Even on that home ground the lead is not universal: Together beat Sarvam on DeepSeek V4.1 Flash with 10,000-token prompts and tied it on GLM 5.3. None of these numbers have been checked independently.
Sarvam also says it runs India's largest cluster of Nvidia B200 GPUs and serves more than 450 billion tokens a day on DeepSeek V4.1 Flash alone, with 95.7% of prompt tokens coming from its cache. Those figures are also unverified.
The local angle may matter more than the benchmarks for Indian businesses. Sarvam says prompts run on its GPUs in India and are not sent to foreign model providers. It offers zero data retention on request, a 99.95% uptime guarantee for enterprise contracts, and ISO 27001 and SOC 2 Type II certification. It also runs the same stack on-premises for what it calls critical institutions.
Comments