Tuesday 17 November 2026Updated 18:04 IST
bharat.digital

India's digital economy, public infrastructure and security

Originals · Startup · Startups & Fintech

Sarvam opens Token Factory to all, serving open AI models from GPUs in India

The Bengaluru AI unicorn now sells inference on its own Sarvam 105B and on open-weight models like DeepSeek, GLM, Kimi and Gemma through one API, billed in rupees and processed in India.

Sarvam opens Token Factory to all, serving open AI models from GPUs in India
Cover: bharat.digital

Sarvam AI has made Token Factory, its inference service, available to all developers and enterprises. The company announced the general launch on its blog on 7 November 2026, and The Economic Times reported it on 16 November. Token Factory serves Sarvam's own models and leading open-weight models from GPU clusters located in India, through a single OpenAI-compatible API.

The launch catalogue includes Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Pro, Kimi K3, Kimi K3.1 Pro and Google's Gemma 4 31B. Sarvam's documentation lists introductory pricing of Rs 10 per million input tokens and Rs 25 per million output tokens for Sarvam 105B, and Rs 14 and Rs 55 for DeepSeek V4.1 Flash, running until the end of 2026. Developers use their existing Sarvam API key, and switching models is a change to one field.

Sarvam says it matches or beats leading global inference providers on speed and cost for most of the models it serves. Its own figures include a first token in 0.2 seconds and 330 tokens per second for a single user on Sarvam 105B, and 95.7% of prompt tokens served from cache on DeepSeek V4.1 Flash, which it says handles more than 450 billion tokens a day. The company also says it runs India's largest B200 cluster. These are vendor-reported numbers, and AlphaSignal's coverage notes that teams should test them against their own workloads.

The pitch to banks, government bodies and regulated firms is data residency. Sarvam says requests run on its own GPUs in India and are not forwarded to the original model developer's overseas API. Zero data retention is available on request, the managed service carries a 99.95% uptime SLA, and the same stack can be deployed on premises for institutions that need it.

Token Factory grows out of Sarvam Inference, which the company launched at its Epoch 2026 developer conference in Bengaluru on 30 July with Sarvam 105B, GLM 5.2 and Gemma 4, Inc42 reported at the time. Cofounder Vivek Raghavan framed it then as a push for token sovereignty: serving a larger share of the AI tokens India consumes from domestic infrastructure rather than overseas clouds.

More originals