
Sarvam Token Factory, a new AI service from the Bengaluru company Sarvam, went on general sale on 16 November 2026. It runs Sarvam's own Sarvam 105B and well-known open-weight models such as DeepSeek, Kimi, GLM and Gemma on GPU clusters located in India, through one OpenAI-compatible API, with prices listed in Indian rupees. For an Australian business with Indian customers, staff or software partners, it is a new option for where AI work is processed, with questions to ask before anyone plugs it in.
What happened
In its post "Introducing Sarvam Token Factory", dated 16 November 2026, Sarvam said it was launching the service "for everyone". It describes Token Factory as its inference stack. Inference is the work of running a trained model to answer requests. Sarvam says the stack serves its own models and open-weight models from GPU clusters located in India.
Open-weight means the model's trained files are published, so companies other than the original maker can host it. That is why one Indian provider can offer models from several labs: alongside Sarvam 105B, the post lists GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3.1 Pro and Gemma 4 31B. All of them sit behind one OpenAI-compatible API, so software written for OpenAI's request format can often be pointed at it with small changes.
Under "Availability", Sarvam says the service is available to developers and enterprises, that it will add models and compute as demand grows, and that teams can choose models "without lock-in". It also offers the same stack on-premise, inside a customer's own data centre.
Key details
The figures below come from Sarvam's Token Factory pricing page and its launch post. Prices are in INR per 1M tokens and exclude GST. A token is a small chunk of text; AI services charge for the tokens you send (input), the tokens they write back (output) and, at a lower rate, repeated text they have already stored (cached input).
| Model on Token Factory | Price (INR per 1M tokens) | How it is sold |
|---|---|---|
| Sarvam 105B (introductory pricing) | ₹10.00 input, ₹0.25 cached input, ₹25.00 output | Serverless, 128K context |
| DeepSeek V4.1 Flash 1017, Standard (introductory pricing) | ₹14 input, ₹55 output | Serverless, for background and batch work |
| DeepSeek V4.1 Flash 1017, Interactive | ₹25 input, ₹105 output | About 200 tokens per second per user, built for coding agents |
| Gemma 4 31B | ₹35.00 input, ₹10.00 cached input, ₹90.00 output | Serverless, 256K context |
| Kimi K3.1 Pro | ₹200.00 input, ₹20.00 cached input, ₹950.00 output | Provisioned throughput, 1M context |
- Serverless and provisioned throughput. Serverless means you pay per token on shared machines. Provisioned throughput means capacity is set aside for you.
- Currency. The page can show US dollars, converted at ₹94 per US dollar. There is no Australian dollar view.
- Reasoning costs. Reasoning tokens, the model's working-out on harder questions, are billed at the output rate.
- Speed claims. Sarvam says Sarvam 105B returns its first token in 0.2 seconds with single-user throughput of 330 tokens per second, and that Gemma 4 31B, measured from Mumbai on a 10K-token prompt, returns its first token in 0.40 seconds, 5x faster than the only other provider serving it. These are Sarvam's own measurements.
Why it matters
Most AI tools used by Australian businesses send text to data centres run by a handful of US and Chinese companies. Sarvam is selling location as a feature.
That matters most for businesses whose work already crosses into India. Sarvam's Sarvam 105B model page says it is built for chat and tool-calling agents in Indian languages and English, and handles native-script, romanised and code-mixed text, where a message switches between an Indian language and English in one sentence. We have covered India's EKA-1 and Dhīr models built to run offline; Token Factory is the commercial, pay-per-use side of the same push.
It also changes the shopping list for open models. We wrote about the DeepSeek V4.1 Flash 1017 upgrade and the release of Kimi K3 open weights. You can now reach versions of those model families through an Indian provider as well as the original makers and other hosts. Each host has its own terms, logging and support, so "same model" does not mean "same service".
Some caution is needed. Sarvam's comparisons with other providers are its own tests. Its post does not say how long prompts are stored, whether they train models, or which countries can sign up. Those answers decide whether it suits your business.
What this means for businesses
If you run an import or export business with Indian buyers, a professional services firm with an offshore team in India, or a support desk that answers customers in Indian languages, this is worth a test, not a switch. Import and export businesses often send this kind of mixed-language text through AI tools.
Start with privacy. If your business is covered by the Privacy Act, APP 8 covers the steps you must take before personal information is disclosed overseas. The OAIC's guidance on commercially available AI products says privacy obligations apply to personal information put into an AI system and to output that contains it, and that a business adopting such a product should do due diligence: whether it has been tested for your intended use, how human oversight is built in, the privacy and security risks, and who will have access to the information. Sending personal information to a provider in India can count as disclosing it overseas.
A short checklist before anyone in your team, or your software provider, connects Token Factory:
- Ask Sarvam in writing where prompts and outputs are stored, for how long, who can access them and whether they are used for training.
- Confirm you can sign up and pay from Australia, and work out what an INR price plus GST means on your monthly bill.
- Test the model on your own work. Sarvam's model page itself advises evaluating the languages, scripts and terms your customers use rather than relying only on benchmark scores.
- Keep the switch easy. Sarvam's API documentation says the endpoint accepts the Bearer key style used by OpenAI-compatible tooling, which helps, but it has also retired an older model, Sarvam-M (24B). Build so a model can be replaced without rewriting your system.
- Update your privacy policy if you start disclosing personal information to an overseas AI provider.
- If you recruit from India, remember a model that reads code-mixed text does not confirm anyone's work history; a candidate verification platform such as Talent Tabloid is built for that check.
- Give AI agents the smallest set of permissions that works; our guide to AI agent permissions, data access and cost covers the basics.
If an Indian-language model could help your quoting, support or document work, the hard part is usually the connection to your existing systems. Our business systems and integrations work and IT consulting cover choosing a provider, testing it on your data and wiring it in so you can change models later. Ask us for a free first consultation and a written quote before you commit.
Frequently asked questions
What is Sarvam Token Factory?
It is a paid API service from Sarvam, a Bengaluru AI company, that runs AI models on GPU clusters located in India. It serves Sarvam 105B plus open-weight models such as DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro and Gemma 4 31B through one API.
Can an Australian business sign up for Sarvam Token Factory?
Sarvam says the service is available to developers and enterprises from 16 November 2026, and prices are listed in INR excluding GST. Its post does not set out country limits, billing outside India or support hours, so ask Sarvam those questions in writing before you plan around it.
Does Sarvam Token Factory keep my data in India?
Sarvam says the models are served from GPU clusters located in India. That describes where the processing happens. It does not by itself tell you how long prompts are kept, who can see them or whether they are used for training, so read the terms and ask before sending customer data.
Is Sarvam Token Factory cheaper than the AI service we use now?
It might be for some workloads, but the comparison depends on your mix of input, cached and output tokens. Sarvam's speed and cost comparisons are its own measurements. Run the same real tasks through both and compare the bills and the answers.
Do Australian privacy rules apply if we send customer data to an Indian AI service?
They can, if your business is covered by the Privacy Act. APP 8 covers the steps you must take before personal information is disclosed overseas, and the OAIC says privacy obligations apply to personal information put into an AI system and to its output.
Key takeaways
- Sarvam Token Factory launched on 16 November 2026 and runs Sarvam 105B and open-weight models from GPU clusters in India through one OpenAI-compatible API.
- Prices are in INR per 1M tokens excluding GST; Sarvam 105B is ₹10.00 input and ₹25.00 output at introductory pricing.
- Speed and cost comparisons are Sarvam's own tests: run your real tasks before you trust them.
- Sending personal information to a provider in India can bring in APP 8, so check storage, access and training terms first.
- Build any AI connection so the model and the provider can be swapped later.