Sarvam AI has launched Sarvam Token Factory, a new AI inference platform designed to provide developers and businesses access to powerful artificial intelligence models with improved speed, cost efficiency and infrastructure based in India. The platform serves Sarvam’s own models alongside leading open-weight AI models through a single API, allowing developers to build AI applications without managing their own GPU infrastructure.
Launched on November 7, 2026, Token Factory is built around optimising how AI models generate responses, including faster inference, prompt caching, GPU utilisation and model serving. Sarvam says its infrastructure can match or outperform leading global inference providers on speed and cost for many of the models it serves, positioning the platform as a potential alternative for developers building AI agents, coding tools and enterprise applications.
What Is Sarvam Token Factory?
Sarvam Token Factory is an AI inference infrastructure platform that enables developers to run AI models without independently managing the hardware and serving systems required to generate responses.
Instead of training a model or deploying GPU infrastructure themselves, developers can access supported models through Sarvam’s API. The platform handles model serving, infrastructure scaling, performance optimisation and request processing.
The platform supports Sarvam’s proprietary models and several open-weight models, including GLM, DeepSeek, Kimi and Gemma. Developers can access multiple models through an OpenAI-compatible API, making it easier to integrate AI capabilities into existing applications.
Token Factory is designed for workloads ranging from simple chatbot interactions to complex coding agents that require long prompts, repeated context and multi-step reasoning.
Sarvam Token Factory Supports Multiple AI Models
Sarvam is offering a growing catalogue of AI models through Token Factory, allowing users to select models according to their specific requirements, including reasoning, coding, speed and conversational performance.
| AI Model | Primary Use |
|---|---|
| Sarvam 105B | Reasoning, coding and Indian-language AI |
| GLM 5.3 | Advanced reasoning and software engineering |
| GLM 5.3 Flash | Faster reasoning and agent workloads |
| DeepSeek V4.1 Flash | Coding, reasoning and general AI applications |
| DeepSeek V4 Pro | Advanced AI reasoning |
| Kimi K3 | General-purpose AI workloads |
| Kimi K3.1 Pro | Coding agents and complex tasks |
| Gemma 4 31B | Text and image understanding |
The model catalogue allows developers to switch between different AI systems depending on the workload. This could help businesses avoid relying on one model for every application while experimenting with different performance and cost combinations.
Sarvam Claims Faster AI Inference and Lower Latency
One of the central selling points of Token Factory is its ability to reduce the time AI applications take to begin generating responses.
Sarvam says its infrastructure has been optimised using techniques such as caching, specialised inference kernels, quantisation, GPU scheduling and separating different stages of AI model processing.
According to Sarvam, Gemma 4 31B delivered a time to first token of 0.40 seconds and an output speed of 308 tokens per second in its testing using a 10,000-token prompt from Mumbai.
| Performance Metric | Sarvam Token Factory | Other Provider |
|---|---|---|
| Time to first token | 0.40 seconds | 2.00 seconds |
| Output speed | 308 tokens/sec | 193 tokens/sec |
The measurements were conducted using a client located in AWS Mumbai, according to Sarvam. The company says its testing covered a 72-hour period, with results based on median performance.
Lower latency is particularly important for AI applications that require frequent interactions, including coding agents, chatbots, customer service systems and real-time conversational applications.
DeepSeek V4.1 Flash Pricing Starts at ₹14 Per Million Input Tokens
Sarvam Token Factory is also competing on AI inference pricing. The company announced introductory pricing for DeepSeek V4.1 Flash, with different endpoints designed for interactive applications and background workloads.
| DeepSeek V4.1 Flash Endpoint | Input Price | Output Price |
|---|---|---|
| Standard | ₹14 per million tokens | ₹55 per million tokens |
| Interactive | ₹25 per million tokens | ₹105 per million tokens |
The Standard endpoint is designed for background processing and batch workloads, while the Interactive endpoint targets applications that require faster responses, including coding agents.
The pricing structure could make it easier for startups, developers and businesses to experiment with AI agents without committing to expensive infrastructure or high inference bills.
However, actual application costs will depend on token consumption, the model selected, prompt length, output generation and usage patterns.
Sarvam Builds a Multi-Layer AI Inference Infrastructure
Token Factory is not simply an API connecting developers to AI models. Sarvam says it has developed an inference stack that optimises the entire process of generating AI responses.
The infrastructure consists of multiple components that work together to reduce unnecessary computation and improve throughput.
Cache-Aware Routing
The system routes requests based on where reusable prompt data is available. This helps avoid repeatedly processing the same information when users return to an AI agent conversation.
Separate Prefill and Decode Processing
AI model inference involves two major stages. Prefill processes the input prompt, while decode generates the response tokens.
Sarvam separates these workloads into different GPU pools to improve resource utilisation and prevent long prompts from slowing down ongoing response generation.
Multi-Tier Caching
Token Factory uses GPU memory and a shared cache infrastructure to store reusable prompt information.
This is particularly useful for coding agents and AI workflows that repeatedly process large amounts of context, including repository files, conversation history and tool outputs.
Sarvam says its DeepSeek V4.1 Flash deployment served more than 450 billion tokens per day across workloads, with 95.7% of prompt tokens served from cache during the reported measurement period.
These optimisations could reduce repeated computation, improve response times and lower infrastructure costs.
Token Factory Targets AI Agents and Enterprise Workloads
Sarvam is positioning Token Factory for more than traditional AI chat applications. The infrastructure is designed to support coding agents, tool-calling agents, document processing, long-context applications and multi-step AI workflows.
The platform supports a context window of up to one million tokens on selected models, allowing developers to process larger documents, codebases and conversation histories in a single workflow.
AI agents can also connect with external APIs, business systems and tools to perform tasks rather than simply generate text.
These capabilities could be particularly valuable for startups developing software agents, enterprise automation systems, AI-powered research tools and business applications that need to process large volumes of information.
Why Running AI Infrastructure in India Matters
Token Factory’s infrastructure is located in India, a strategic advantage for businesses concerned about latency, data residency and the location where AI requests are processed.
Sarvam says inference requests run on its GPUs in India rather than being forwarded to overseas model providers. The company also offers zero data retention arrangements on request, subject to the applicable deployment terms.
Local infrastructure could become increasingly important for Indian enterprises, financial institutions, government-linked organisations and businesses that have specific data-handling requirements.
It could also improve access to AI infrastructure for Indian startups that want to build AI products without depending entirely on international cloud providers.
Sarvam Token Factory Could Intensify AI Infrastructure Competition
The launch comes as AI infrastructure providers compete to make model inference faster and cheaper. While model training attracts significant attention, inference is the infrastructure required every time a user asks an AI system a question or an AI agent performs a task.
As AI applications gain users, inference costs can become a major operational expense. Faster processing, higher GPU utilisation and caching can help providers serve more requests using the same infrastructure.
Sarvam’s strategy focuses on optimising these underlying systems instead of relying exclusively on the capabilities of individual AI models.
By combining multiple models, local GPU infrastructure, a unified API and inference optimisations, the company is attempting to compete with international AI infrastructure platforms on performance, price and accessibility.
The Bigger Picture
Sarvam Token Factory reflects India’s growing ambition to build AI infrastructure rather than only consume AI services developed elsewhere. The platform gives developers access to multiple models through infrastructure operated in India, potentially helping businesses reduce latency, improve data control and manage AI deployment costs.
The bigger opportunity is the growing demand for AI agents capable of completing tasks, writing code, processing documents and interacting with business systems. These applications require dependable inference infrastructure capable of handling large volumes of requests without dramatically increasing costs.
Looking Ahead
Sarvam plans to expand the models available through Token Factory and increase its computing capacity as demand grows. The company is also developing additional tools for developers to build, deploy and scale AI agents on its infrastructure. The platform’s ability to maintain performance as usage increases will be important as businesses move AI projects from experimentation to production.
As the AI infrastructure market becomes more competitive, pricing, response speed, model availability, reliability and data privacy will increasingly determine which platforms developers choose. Sarvam Token Factory could strengthen India’s position in the global AI ecosystem if it can consistently deliver competitive performance while supporting the needs of startups and large enterprises.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



