Editorial illustration for Sarvam launches Token Factory for India-hosted inference on 8 open models
Sarvam Token Factory Launches India Inference Platform
Sarvam has taken its inference platform out of preview and opened it to any developer who wants it. Sarvam Token Factory now serves eight models, including the company's own Sarvam 105B alongside open-weight options like GLM 5.3, DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro and Gemma 4 31B, all running on GPU clusters based in India. The pitch is straightforward: teams building in India can get tokens back fast without routing requests through servers in another country. One OpenAI-compatible API sits in front of all eight models, so switching between them doesn't mean rewriting integration code.
The platform first surfaced under a different name, Sarvam Inference, when the company previewed it at its Epoch event. Since then it's been restructured into two tiers. Some models run on shared, pay-as-you-go endpoints.
Others, including DeepSeek V4 Pro and Kimi K3.1 Pro, require Provisioned Throughput, a setup enabled per organization rather than offered on demand. Sarvam has also published its own benchmark numbers comparing throughput and latency against at least one rival provider.
Sarvam Token Factory is an inference platform from Sarvam that serves its own models and open-weight models from GPU clusters in India. Developers reach every model through one OpenAI-compatible API. For India-based teams, this means low-latency tokens without sending traffic to overseas clouds.
Why this matters
For developers building in or for India, Token Factory removes a real friction point: the choice between latency and data residency. Sarvam's own numbers, 0.2s first token on its 105B model, 0.40s versus 2.00s for Gemma 4 31B on a 10K prompt out of Mumbai, aren't neutral third-party benchmarks, so treat them as a vendor's best-case showcase rather than gospel. Still, the gap against "the only other provider" they cite is wide enough to be worth testing yourself before committing infrastructure.
The bigger signal is strategic: Sarvam is positioning itself as the default inference layer for India-based teams who don't want to route traffic through US or Singapore clouds for compliance or speed reasons. One OpenAI-compatible API across 8 models, including third-party ones like Kimi K3.1 Pro, lowers the switching cost for teams currently stitching together multiple providers. If the 330 tokens/s and sub-second latency numbers hold up under independent load, this becomes a legitimate regional alternative rather than a nationalist talking point. Worth watching whether other India-focused labs follow with their own hosted inference layers, or whether this stays a one-player market.
Common Questions Answered
What models are available through Sarvam Token Factory?
Sarvam Token Factory serves eight models including Sarvam's own Sarvam 105B alongside open-weight options such as GLM 5.3, DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro, and Gemma 4 31B. All of these models run on GPU clusters physically located in India, allowing developers to access them through a single OpenAI-compatible API.
What is the main advantage of using Sarvam Token Factory for India-based developers?
The primary advantage is achieving low-latency token generation without routing requests through overseas servers, which removes the friction point between latency and data residency concerns. Developers building in or for India can now get tokens back quickly while keeping traffic within the country's infrastructure.
How does Sarvam Token Factory's latency compare to other providers?
According to Sarvam's benchmarks, the platform achieves 0.2 seconds for first token on its 105B model and 0.40 seconds for Gemma 4 31B on a 10K prompt from Mumbai, compared to 2.00 seconds for the same task with another provider. However, these are vendor-provided benchmarks rather than independent third-party measurements, so they should be tested independently before making commitments.
What is the OpenAI-compatible API in Sarvam Token Factory?
The OpenAI-compatible API is a unified interface that allows developers to access all eight models available on Token Factory through a single standardized API endpoint. This compatibility means developers familiar with OpenAI's API can easily integrate Sarvam's inference platform without learning a new API structure.
Why did Sarvam transition Token Factory from preview to general availability?
Sarvam opened Token Factory to any developer to provide India-based teams with a practical solution for low-latency inference without the need to route requests to international servers. This move addresses a specific friction point for developers building in India who need both fast performance and data residency compliance.
Further Reading
- Sarvam AI launches Token Factory to offer cheaper AI model inference from India - The Economic Times
- Sarvam opens AI token factory, India’s answer to OpenRouter and Hugging Face - India Today
- Introducing Sarvam Token Factory - Sarvam AI
- Open-sourcing Sarvam 30B & 105B models - Sarvam AI