Sarvam AI's Token Factory is a paid inference API that lets you call several open-weight language models, including Chinese ones from DeepSeek, Moonshot AI (Kimi) and Z.ai (GLM), through one OpenAI-compatible endpoint that Sarvam says runs on GPUs in India. It launched on 16 November 2026 and immediately drew criticism because a company seen as the face of India's sovereign-AI effort was selling access to models built abroad. This guide explains what the service offers and costs, what each side of the debate says, and how to judge any model host on your own data and compliance needs.
Who this is for
This is for developers, IT leads and founders in India who are deciding whether an India-hosted LLM API is worth using. It is not a verdict on Sarvam or on the politics around it. By the end you will know:
- what Token Factory is and how to call it,
- what the listed models cost, with a worked monthly estimate,
- what critics and Sarvam each said, attributed to the people who said it,
- which questions to ask before you send personal data to any LLM provider, including Sarvam.
What Token Factory is
Sarvam's Token Factory documentation says the service "serves leading open-weight LLMs on Sarvam infrastructure through one OpenAI-compatible API." It shares the API key, SDK and credits of Sarvam's other APIs, so there is no separate signup. Sarvam's own site describes the platform as developed and operated in India and lists three deployment options: Sarvam Cloud (managed), Private Cloud (VPC) and on-premises.
In practice that means a team already using the OpenAI SDK can change two things, the base URL and the key, and then pick a model by name.
| Item | Value from Sarvam's docs |
|---|---|
| Base URL | https://api.sarvam.ai/v2 |
| Chat endpoint | POST /v2/chat/completions |
| Embeddings endpoint | POST /v2/embeddings |
| Model selection | the model field, for example glm5.3-flash |
| Auth in the cURL example | api-subscription-key header |
A minimal call with the OpenAI Python SDK looks like this. The base URL and model ID come from Sarvam's documentation; the key is a placeholder.
from openai import OpenAI
client = OpenAI(
base_url="https://api.sarvam.ai/v2",
api_key="YOUR_SARVAM_API_KEY",
)
response = client.chat.completions.create(
model="glm5.3-flash",
messages=[{"role": "user", "content": "Summarize this ticket in two lines."}],
)
print(response.choices[0].message.content)Because the model is just a field, you can run the same prompt against several models and compare quality and cost on your own data, which is the only comparison that matters for your workload. For help with counting tokens before you budget, see LLM token counting and context window limits.
Which models are listed
Sarvam's pricing page separates models that are serverless from those on Provisioned Throughput, which is enabled per organization. The list below is what the page showed on 19 November 2026.
| Model | Model ID | Developer | Availability |
|---|---|---|---|
| Sarvam 105B | sarvam-105b | Sarvam | Serverless (introductory price) |
| Sarvam 105B Chat | sarvam-105b-conversations | Sarvam | Serverless (introductory price) |
| GLM-5.3 | glm5.3 | Z.ai | Serverless |
| GLM-5.3 Flash | glm5.3-flash | Z.ai | Serverless |
| Gemma 4 31B | gemma4 | Serverless | |
| DeepSeek V4.1 Flash 1017 | deepseekv4.1-flash-1017 | DeepSeek | Serverless (introductory price) |
| DeepSeek V4.1 Flash 1017 Interactive | deepseekv4.1-flash-1017-interactive | DeepSeek | Serverless |
| DeepSeek V4 Pro | deepseekv4-pro | DeepSeek | Provisioned Throughput |
| Kimi K3.1 Pro | kimi-k3.1-pro | Moonshot AI | Provisioned Throughput |
| Qwen 3.8 27B | qwen38-27b | Alibaba (Qwen) | Provisioned Throughput |
The page also lists Sarvam Prism, a model that returns probabilities rather than text, plus Clef and EmbeddingGemma 2, both billed on input tokens only. Developer attribution in the table is our reading of the model names, not a column on Sarvam's page.
News outlets count the catalog differently. Analytics India Magazine and MediaNama (as summarized in search results) report eight models, six of them Chinese. FourWeekMBA counts 12 listings on the price page. The gap is most likely serverless versus provisioned entries, but that is our inference, not something the sources state.
What it costs
All prices on Sarvam's pricing page are in rupees per million tokens, exclude GST and are the same on every plan. Reasoning tokens, where a model produces them, are billed as output tokens. If credits run out, requests fail with a 402 error, which is separate from a 429 rate-limit error.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Sarvam 105B | Rs 10 | Rs 0.25 | Rs 25 |
| DeepSeek V4.1 Flash 1017 | Rs 14 | Rs 0.25 | Rs 55 |
| GLM-5.3 Flash | Rs 13.5 | Rs 2.5 | Rs 45 |
| Gemma 4 31B | Rs 35 | Rs 10 | Rs 90 |
| GLM-5.3 | Rs 126 | Rs 20 | Rs 390 |
Introductory pricing for Sarvam 105B, its Chat variant and the standard DeepSeek V4.1 Flash 1017 runs through the end of 2026. After that, plan for a different rate and check the page again.
Conflicting prices in the news
Not every outlet matches the pricing page. Inc42 reported Sarvam 105B at Rs 15 input and Rs 60 output with cached input at Rs 5. Sarvam's pricing page and Token Factory page both showed Rs 10 and Rs 25 when we opened them. The most likely explanation is that prices changed or that the outlet quoted a different tier, but we could not confirm which. Use the vendor page on the day you buy and keep a screenshot for your records.
A worked monthly estimate
Take a support-summarization job that uses 50 million input tokens and 10 million output tokens a month, with no caching. The cost is input millions times the input price plus output millions times the output price.
| Model | Calculation | Monthly cost, excluding GST |
|---|---|---|
| Sarvam 105B | 50 x 10 + 10 x 25 | Rs 750 |
| GLM-5.3 Flash | 50 x 13.5 + 10 x 45 | Rs 1,125 |
| DeepSeek V4.1 Flash 1017 | 50 x 14 + 10 x 55 | Rs 1,250 |
| Gemma 4 31B | 50 x 35 + 10 x 90 | Rs 2,650 |
| GLM-5.3 | 50 x 126 + 10 x 390 | Rs 10,200 |
Cheaper is not automatically better. A smaller model that needs two retries costs more than a larger one that answers correctly the first time, and reasoning tokens raise the output side. Run your own test set before committing.
The debate, with attribution
This is a fast-moving story and most of the reporting is a day or two old. The points below are what each outlet or person said, not our assessment.
The criticism
- Analytics India Magazine (18 November 2026) noted that six of the eight listed models come from Chinese companies and only one is Sarvam's own, calling the optics hard to ignore.
- Critics cited in the reporting pointed to Sarvam's role in India's sovereign-AI push. Analytics India Magazine reports a Rs 98.68 crore IndiaAI Mission GPU subsidy covering 4,096 NVIDIA H100 GPUs, and that Sarvam raised $234 million at a $1.5 billion valuation in June.
- Forbes contributor Janakiram MSV argued that pricing Sarvam's own model below DeepSeek turns "token sovereignty" into a commercial pitch, and raised the risk that developers come for DeepSeek, GLM or Kimi rather than Sarvam's model. We read this from a search summary and did not open the full article.
Sarvam's response
- Co-founder Pratyush Kumar, quoted by Careerindia, said: "The token factory is running on GPUs privately procured by Sarvam." Careerindia and MediaNama's summary also report that subsidized IndiaAI compute ended in March 2026.
- Sarvam says prompts are processed in India and are not forwarded to model developers' overseas APIs (Inc42, Careerindia). Careerindia adds that Sarvam offers optional zero data retention for enterprises. Both are Sarvam's claims, reported without independent verification.
- Lightspeed's Hemant Mohapatra, an investor in Sarvam who said he was biased, wrote on X that global inference providers earn significant revenue hosting global models, including Chinese ones, in India (Analytics India Magazine).
- MediaNama's founder Nikhil Pahwa was reported as saying he sees no problem with Sarvam running an inference platform or hosting Chinese models.
What is still unverified
Sarvam's speed and cost claims, such as matching or beating global providers on most hosted models, come from Sarvam's own tests. Analytics India Magazine reports that a competing Indian provider says its benchmarks are on par or faster on three models, which is also a self-interested claim. Neither side has been independently verified.
The launch date also differs by source: some reports are dated 17 or 18 November. This guide uses 16 November, the date on Sarvam's own launch post and in Analytics India Magazine.
Hosted in India is not the same as safe
Two questions get mixed together in this debate, and it helps to separate them.
- Where is my data processed? A model hosted in India and run on the provider's own GPUs avoids sending prompts to a model developer's overseas API. That is a data-residency question.
- What does the model do and who sees the data? The weights come from a developer abroad, but open weights running on someone else's hardware do not phone home by themselves. What matters is the host's logging, retention, access controls and contract, and the model's own quality and behavior, including its safety tuning and any topics it handles differently.
Residency answers the first question and says nothing about the second. A model host in Mumbai with unlimited log retention can be a worse choice for personal data than a foreign host with a strict zero-retention contract.
DPDP Act points to check first
India's Digital Personal Data Protection Act, 2023 is relevant whenever you send personal data to an LLM API. The Rules were notified in November 2025 with a phased rollout: law-firm summaries report three stages, 14 November 2025, 14 November 2026 and 14 May 2027. Confirm which obligations apply to you on the date you deploy by reading the MeitY notification.
The Act text (as reproduced on dpdpa.com) says:
- Section 8(1): a Data Fiduciary is "responsible for complying with the provisions of this Act and the rules made thereunder in respect of any processing undertaken by it or on its behalf by a Data Processor." If you are the company deciding why the data is processed, you remain responsible when an API handles it.
- Section 8(2): a Data Processor may be engaged "only under a valid contract."
- Section 8(5): you must take "reasonable security safeguards to prevent personal data breach," including for processing done on your behalf.
- Section 8(7)(b): you must "cause its Data Processor to erase any personal data that was made available" to it when erasure is due, which means you need to know how to make your LLM host delete data.
- Section 16(1): the Central Government "may, by notification, restrict the transfer of personal data" to a country outside India. Section 16(2) preserves stricter sector laws, which matters if you are a bank or insurer.
This is a summary for orientation, not legal advice. Have counsel review your contract with any AI provider. For the website side of the same law (privacy notices, consent banners, rights requests and breach timers), see the guide to the open-source EpiFive Website Privacy Kit.
How to evaluate any model host
Use the same checklist for Sarvam, a hyperscaler region, or any other provider. Write the answers down and keep the source for each.
| Question | What a good answer looks like |
|---|---|
| Where are prompts processed? | A named region, stated in the contract or a trust page, not only in a news report |
| Is anything logged or retained? | A stated retention period, and a zero-retention option if you need it |
| Is data used for training? | An explicit "no" in the terms |
| Who are the sub-processors? | A list, including any hosting or observability vendors |
| What compliance reports exist? | Current audit reports you can request. Sarvam's page lists SOC 2 and ISO 27001 badges, and its site was inconsistent about the SOC 2 type, so ask for the report |
| What are the service terms? | An uptime commitment, support channel and incident process |
| How do pricing and limits work? | Published rates, rate limits and a clear notice period for price changes |
| Can you leave? | A standard API such as the OpenAI format, so switching is a base URL change |
Sarvam's site claims a 99.9% uptime SLA and data residency controls, and Sarvam's Token Factory docs page contains no retention or residency statements. That is a documentation gap to close with the vendor, not proof of a problem.
A short evaluation routine
- Build 30 to 50 real prompts from your own workload, with personal data removed or replaced.
- Send the same set to two or three models by changing the
modelfield. - Score answers for correctness and format, and record tokens used per call.
- Compute the cost for your volume using the formula above.
- Only then decide whether the residency and compliance answers justify the price.
Alternatives, without a ranking
Inc42 lists the competitive field: Together AI, Fireworks AI, OpenRouter and the major clouds (AWS, Microsoft Azure, Google Cloud). It also reports that Krutrim has hosted DeepSeek models on Indian servers since January 2025, that OpenAI introduced India-based inference for eligible workloads on 12 November, and that Anthropic offers in-country inference for select Claude models through Amazon Bedrock. We did not verify those last two on the vendors' own pages, so check each vendor's documentation for which models, regions and plans actually qualify.
Troubleshooting
- 402 error: Sarvam's pricing page says this means credits ran out. Add credits and retry. A 429 is a separate rate limit.
- Model not found: provisioned models (DeepSeek V4 Pro, Kimi K3.1 Pro, Qwen 3.8 27B) are enabled per organization, so an account without access should not expect them to work. Use a serverless model ID or ask Sarvam to enable access.
- Costs higher than the estimate: reasoning tokens count as output tokens, and long system prompts count on every call. Log token usage per request.
See also
- Azure API Management AI Gateway: Token Limits and Load Balancing for OpenAI
- Claude API Pricing: How to Estimate Token Costs for a Business Workflow
- LLM Model Routing: Send Easy Requests to a Cheaper Model
- How to Set Up Gemini API Tool Retrieval: Models, Pricing, Limits
What we checked
- Opened Sarvam's Token Factory docs page on 2026-11-19. It describes an OpenAI-compatible API at
https://api.sarvam.ai/v2and lists the model IDs in this article. It states nothing on retention, logging, residency or rate limits. - Opened Sarvam's pricing page on 2026-11-19 and copied the rupee prices, cached prices and availability labels in the tables above, including the end-of-2026 introductory pricing note and the 402 versus 429 billing note.
- Cross-checked that page against the Token Factory docs: DeepSeek V4.1 Flash 1017 (Rs 14 and 55), Sarvam 105B (Rs 10 and 25), GLM-5.3 (Rs 126 and 390), GLM-5.3 Flash (Rs 13.5 and 45) and Gemma 4 31B (Rs 35 and 90) matched. Inc42's Rs 15 and Rs 60 for Sarvam 105B did not match, as noted above.
- Recalculated the monthly estimate with 50 million input and 10 million output tokens: Sarvam 105B 500 + 250 = Rs 750; GLM-5.3 Flash 675 + 450 = Rs 1,125; DeepSeek V4.1 Flash 1017 700 + 550 = Rs 1,250; Gemma 4 31B 1,750 + 900 = Rs 2,650; GLM-5.3 6,300 + 3,900 = Rs 10,200.
- Read the Inc42 (17 November), Analytics India Magazine (18 November) and Careerindia reports and attributed each claim to its source. MediaNama's and Forbes's pages returned errors or were read only as search summaries, so those points are marked as such.
- Read Section 8 and Section 16 of the DPDP Act as reproduced on dpdpa.com. The Gazette and indiacode.nic.in copies returned HTTP 403, so the wording was not compared with the official PDF.
- Not checked: we made no API calls, so the Python snippet and model behavior are untested, and we did not verify Sarvam's speed, uptime, residency or retention claims or the DPDP Rules timeline against MeitY.
Last tested: 2026-11-19
References
- Sarvam Token Factory docs: https://docs.sarvam.ai/api/token-factory
- Sarvam API pricing: https://docs.sarvam.ai/api/pricing
- Sarvam: https://www.sarvam.ai/
- Inc42, Sarvam AI opens Token Factory: https://inc42.com/buzz/sarvam-ai-opens-token-factory-to-run-global-ai-models-on-indian-servers/
- Analytics India Magazine, Token Factory sparks sovereign AI debate: https://analyticsindiamag.com/ai-features/sarvams-token-factory-sparks-indias-sovereign-ai-debate-again
- Careerindia, Sarvam Token Factory: https://www.careerindia.com/news/sarvam-token-factory-chinese-ai-models-sovereignty-concerns-2026-011-68881.html
- MediaNama: https://www.medianama.com/2026/11/223-sarvam-token-factory-india-sovereign-ai/
- Forbes: https://www.forbes.com/sites/janakirammsv/2026/11/19/sarvam-prices-its-own-model-below-deepseek-in-new-token-factory/
- DPDP Act 2023, Section 8: https://dpdpa.com/dpdpa2023/chapter-2/section8.html
- DPDP Act 2023, Section 16: https://www.dpdpa.com/dpdpa2023/chapter-4/section16.html
Comments
Questions and tips are welcome. Comments are checked by the EpiFive team before they appear; please do not share passwords, phone numbers or other private details.