Skip to content
PressPire PressPire PressPire

Stories that shape the future

PressPire PressPire PressPire

Stories that shape the future

  • Home
  • Founder Stories
  • Startup Stories
  • News
    • Funding News
    • Trending News
  • Women In Spotlight
  • Industry Verticals
    • Business
    • Education
    • Lifestyle
    • Healthcare
    • Legal
    • Technology
    • Thought Leadership
    • Others
  • Podcasts
  • Home
  • Founder Stories
  • Startup Stories
  • News
    • Funding News
    • Trending News
  • Women In Spotlight
  • Industry Verticals
    • Business
    • Education
    • Lifestyle
    • Healthcare
    • Legal
    • Technology
    • Thought Leadership
    • Others
  • Podcasts
  • Linkedin
  • Instagram
  • Facebook
  • Youtube
Home/Industry Verticals/Technology/Sarvam Token Factory brings Indian and global AI models to Indian infrastructure
Sarvam Token Factory
Technology

Sarvam Token Factory brings Indian and global AI models to Indian infrastructure

5 Min Read

The platform gives developers and enterprises access to multiple AI models through a shared interface, with inference processed on GPU infrastructure in India.

Sarvam Token Factory is now available to developers and enterprises, bringing Sarvam’s own AI models and selected global open-weight models to servers in India. In its launch announcement dated 16 November 2026, the company positioned the service around model choice, performance and the location of AI processing.

The launch extends Sarvam’s offering beyond its own models. Businesses can access models from different developers while using a common service for hosting and inference.

For customers, the practical question is how well that combination works in production: whether the chosen model provides useful answers, responds quickly enough and fits the application’s budget and data-handling requirements.

What Sarvam Token Factory does

Token Factory is an AI inference service. Inference is the process of running a trained model to generate an answer, analyse information or perform another supported task.

A developer connects an application to the service through an application programming interface, or API. The application sends a request, the selected model processes it and the response returns to the application.

Sarvam’s documentation describes an OpenAI-compatible interface that uses the same API key and credits as its other APIs. Developers can select supported models through the request’s model setting.

That shared interface can reduce integration work. Applications still need testing when models change because their supported features, response formats and behaviour can differ.

Indian and global models in one catalogue

The model selection on Sarvam Token Factory includes Sarvam 105B alongside models from the GLM, DeepSeek, Gemma, Kimi and Qwen families.

Examples in the current catalogue include GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Gemma 4 31B and Kimi K3.1 Pro. Access arrangements vary by model.

Open-weight models make their trained parameters available under specified licences, allowing them to be hosted by providers such as Sarvam. The applicable licence still governs their use.

Sarvam’s own 105B model adds an Indian-language component to the offering. Its documentation lists support for ten widely spoken Indian languages and English, including native-script, romanised and mixed-language input. A conversational variant is designed for dialogue and voice-agent workloads.

For a developer, the useful outcome is a choice of models that can be evaluated against the actual task, whether that involves customer conversations, coding or document analysis.

How Sarvam Token Factory pricing works

The service charges for model usage in tokens: the pieces of text a model receives or generates. Input, cached input and output can have different rates.

Selected published prices, checked on 20 November 2026, are:

ModelInput price per million tokensOutput price per million tokens
Sarvam 105B — introductory₹10₹25
DeepSeek V4.1 Flash 1017 Standard — introductory₹14₹55
GLM 5.3 Flash₹13.50₹45
Gemma 4 31B₹35₹90

These rates exclude GST. Sarvam’s documentation states that the listed introductory offers run through the end of 2026. Reasoning tokens, where applicable, are billed as output.

For example, one million uncached input tokens and one million billable output tokens on Sarvam 105B would cost ₹35 before GST at the introductory rates.

Actual spending depends on the selected model, prompt length, output volume, reasoning usage and eligibility for cached-input pricing. A production budget therefore needs to reflect the application’s typical requests.

What Indian hosting means for data handling

A central feature of Sarvam Token Factory is the location of model execution. Sarvam says it hosts the model weights on its own operated GPUs in India and does not forward inference requests to the original developers’ overseas APIs.

That gives customers a clear location for the inference step. Other services connected to an application, such as external tools, have their own processing locations and data policies.

Retention is another separate setting. Zero data retention is available on request, with prompts and responses covered by an agreed arrangement. Operational-log retention is handled separately.

Sarvam’s current product page also says content may be used for model training or improvement unless users opt out. Its privacy policy states that enterprise customer content is governed by the relevant customer agreement.

These details make deployment terms important. Hosting location, retention and permission to use submitted content each answer a different question about how an application handles data.

Access options for developers and enterprises

Sarvam Token Factory offers both self-service access and models that require capacity to be arranged for an organisation.

Its documentation distinguishes two approaches:

  • Serverless access: Supported models can be called using a Sarvam API key, with usage charged per token.
  • Provisioned throughput: Capacity is configured for an organisation’s workload and access must be enabled.

Models such as DeepSeek V4 Pro and Kimi K3.1 Pro currently use provisioned throughput. Their presence in the catalogue therefore does not mean every account can call them immediately.

Sarvam also says its inference software can be deployed on an enterprise’s premises. This gives organisations another deployment option when they need to manage the infrastructure environment more directly.

Understanding the performance claims

Performance is another part of Sarvam Token Factory’s pitch. In its launch material, Sarvam reported a 0.2-second time to first token and single-user output speed of 330 tokens per second for Sarvam 105B.

The company also describes caching and changes to how processing work is assigned across GPUs. These techniques aim to reduce repeated computation and improve the use of available hardware.

The published figures are company benchmarks. Results in a particular application will depend on prompt length, traffic, model settings and the work being requested.

For customers, an effective comparison should consider answer quality and total task cost alongside response speed.

The business opportunity behind the launch

The business opportunity for Sarvam Token Factory lies in providing a service that applications use repeatedly as their traffic grows.

For a customer, managed inference can reduce the work involved in buying hardware, deploying model weights and maintaining a serving system. That can allow a development team to spend more time on its product.

For Sarvam, the commercial challenge is to earn enough from usage to support computing capacity, engineering and customer service. Efficient hardware use matters, but so does maintaining performance when many customers submit requests at once.

The broader significance for India’s AI industry is practical: domestic infrastructure can support both locally developed models and selected models built elsewhere. The value of that combination will depend on reliable operation and competitive economics.

What will determine long-term adoption

The long-term test for Sarvam Token Factory is whether developers continue using it after evaluation.

Model availability, application quality, predictable billing and dependable service will all influence that decision. Enterprises will also assess whether the deployment arrangements fit their internal requirements.

A useful model catalogue attracts attention. Sustained adoption will depend on helping teams run applications consistently as their workloads, users and model choices change.


For more insights navigate: PressPire | Stories That Shape the Future

Read previous article here : deepak-rajawat-air-india-express-ceo/

Follow PressPire on Facebook | Instagram | LinkedIn | YouTube

PressPire | Stories That Shape the Future

Tags:

AI InferenceData ResidencyIndian AIOpen-Weight ModelsSarvam AIToken Factory
Author

presspireindia@gmail.com

Follow Me
Other Articles
Deepak Rajawat
Previous

Deepak Rajawat appointed Air India Express CEO, takes charge on December 1, 2026

No Comment! Be the first one.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Search

Recent Posts

  • Sarvam Token Factory brings Indian and global AI models to Indian infrastructure
  • Deepak Rajawat appointed Air India Express CEO, takes charge on December 1, 2026
  • Kia Andhra Pradesh plant: $2 billion plan for 12,000 jobs and production in 2030
  • Mappls Arattai integration: MapmyIndia brings navigation to Zoho’s messaging app
  • Banofi: Turning Banana Plant Waste Into a Leather Alternative for Modern Fashion

About PressPire

We help founders, entrepreneurs, executives, investors, creators, and industry leaders transform their expertise into influence through strategic media visibility, thought leadership, podcast storytelling, and personal brand building.

Contact Us

Address
Emperize Universe Media
Office-216, A-building, Gera’s Imperium Gateway, Pune, Maharashtra 411034

Hours
Monday–Friday: 10:00AM–8:00PM

We help founders, entrepreneurs, executives, investors, creators, and industry leaders transform their expertise into influence through strategic media visibility, thought leadership, podcast storytelling, and personal brand building.

Vital Links

  • About Us
  • Podcasts
  • Terms & Conditions
  • Privacy Policy

Contact Us

Address
Emperize Universe Media
Office-216, A-building, Gera’s Imperium Gateway, Pune, Maharashtra 411034

Hours
Monday–Friday: 10:00AM–8:00PM

Copyright 2026 — PressPire. Developed by Quadrang Systems