Sarvam has opened Token Factory to everyone. It is a hosted inference API that serves Sarvam's own 105B model and several open-weight models from other labs on GPUs in India, billed in rupees. Sarvam announced it on November 16, 2026, and says it matches leading inference providers on performance and reliability. The cheapest chat model on the list, Sarvam 105B, costs ₹10 per million input tokens and ₹25 per million output tokens.
Every speed and cache figure below comes from Sarvam, not from an independent test. The prices come from Sarvam's own pricing page, which we read on the launch day. We could not open the signed-in Token Factory console, so we have not run a request through it ourselves.
TL;DR: what Token Factory is, in a table
| Question | Answer |
|---|---|
| What is it? | A managed API for open-weight LLMs, run on Sarvam's GPUs in India |
| Is it open to everyone? | Yes, as of November 16, 2026 |
| API shape | OpenAI-compatible, base URL https://api.sarvam.ai/v2 |
| Cheapest chat model | Sarvam 105B: ₹10 input, ₹0.25 cached input, ₹25 output per 1M tokens |
| Gemma 4 31B | ₹35 input, ₹10 cached, ₹90 output per 1M tokens |
| DeepSeek V4.1 Flash 1017 | ₹14 input, ₹0.25 cached, ₹55 output (introductory, through end of 2026) |
| Currency | Indian rupees |
| Rate limit on pay as you go | 50 requests per minute on gemma4, glm5.3, deepseekv4.1-flash-1017; 60 on Sarvam 105B |
| Who checked the speed claims? | Sarvam only so far |
Which models can you call?
Sarvam says the launch catalogue covers its own models and "the full range of best-in-class open-weight models." The docs list these model IDs.
| Model | Model ID | Access | Notes |
|---|---|---|---|
| Sarvam 105B | sarvam-105b | Serverless | Indian-language reasoning, coding, agents |
| Sarvam 105B Conversations | sarvam-105b-conversations | Serverless | Tuned for real-time dialogue |
| GLM-5.3 | glm5.3 | Serverless | ₹126 in, ₹390 out |
| GLM-5.3 Flash | glm5.3-flash | Serverless | Image input, ₹13.5 in, ₹45 out |
| Gemma 4 31B | gemma4 | Serverless | Images and low-latency answers |
| DeepSeek V4.1 Flash 1017 | deepseekv4.1-flash-1017 | Serverless | Long documents, cost-sensitive reasoning |
| DeepSeek V4.1 Flash 1017 Interactive | deepseekv4.1-flash-1017-interactive | Serverless | ₹25 in, ₹105 out, aimed at coding agents |
| DeepSeek V4 Pro | deepseekv4-pro | Provisioned throughput | Enabled per organisation |
| Kimi K3.1 Pro | kimi-k3.1-pro | Provisioned throughput | ₹200 in, ₹950 out |
| Qwen 3.8 27B | qwen38-27b | Provisioned throughput | Enabled per organisation |
| Sarvam Prism | sarvam-prism | Serverless | Decision model, called through POST /v2/systemone |
| Clef, EmbeddingGemma 2 | clef, embeddinggemma-2 | Provisioned throughput | Input-only billing |
Two notes on that table. The docs list Kimi K3.1 Pro, while the AlphaSignal launch report also lists Kimi K3, so check the live catalogue for the exact set. "Provisioned throughput" means the model is turned on per organisation, so you cannot call it from a fresh pay-as-you-go key.
What does it cost, and is the "80% cheaper" claim right?
Sarvam's launch thread says Sarvam 105B is "80% cheaper than Gemma 4 31B." We ran the arithmetic on the list prices from the pricing page.
| Workload | Sarvam 105B | Gemma 4 31B | Saving |
|---|---|---|---|
| 1M input tokens | ₹10 | ₹35 | 71% |
| 1M output tokens | ₹25 | ₹90 | 72% |
| 1M input + 200K output | ₹15 | ₹53 | 72% |
| 1M cached input | ₹0.25 | ₹10 | 98% |
The saving is large, but on uncached list prices it lands near 72%, not 80%. The thread does not show how Sarvam reached 80%. A workload with heavy cache hits would push the gap higher, because cached input on Sarvam 105B is only ₹0.25. If you rely on that number for a budget, run your own token mix instead.
The price is also an introductory price. Sarvam says introductory pricing on some models runs through the end of 2026. Plan for the rate to change.
What do the plans change?
Language model prices are the same on every plan. The plans change limits and support:
- Pay as you go: no monthly fee, 50 to 60 requests per minute.
- Pro: ₹9,999 per month, 200 requests per minute.
- Growth: ₹49,999 per month, up to 1,000 requests per minute on default models.
- Business: ₹1,99,999 per month, includes ₹1,99,999 of credits.
- Enterprise: custom.
The pricing page does not publish rate limits for every model or for Business and Enterprise.
How fast is it, according to Sarvam?
Sarvam published several speed claims in its thread. They are vendor numbers.
- Sarvam 105B: first token in 200 ms, with single-user throughput of 330 tokens per second. Sarvam says it optimised kernels, trained custom draft models and served the model in prefill and decode disaggregated mode, for a 2.86x throughput gain over its bf16 version.
- Gemma 4 31B: first token 5x sooner and 1.6x more tokens per second than the only other provider serving it. AlphaSignal reports 0.40 s to first token on a 10,000-token prompt and 308 against 193 tokens per second. Sarvam says it already serves 150 billion tokens per day of this model. For a different speed claim on the same model, see our Gemma 4 31B on Cerebras post.
- DeepSeek V4.1 Flash 1017: an Interactive endpoint at 200 tokens per second per user for coding agents, and a cheaper Standard endpoint for background and batch work. For earlier context on this family, read DeepSeek V4 Flash with Codex and the Responses API.
- GLM 5.3 Flash and Kimi K3.1 Pro: up to 2.1x sooner first token on GLM 5.3 Flash than the stock open-source stack, and up to 1.9x more tokens per GPU on Kimi K3.1 Pro against the best open stack Sarvam compared.
A cache fast lane beside a slow road, showing how prefix caching cuts latency on Sarvam Token Factory
Why caching is the real story
The cheapest token is the one you do not recompute. Sarvam says it runs a three-tier cache that keeps context across requests, and that its router and admission layer account for the distributed cache. On DeepSeek V4.1 Flash, 95.7% of prompt tokens are served from cache across more than 450 billion tokens a day.
This matters most for coding agents. Agent prompts repeat the same long prefix on every turn. AlphaSignal reports that GLM 5.3 agent prompts average 131,000 tokens. When the prefix is cached, you pay the cached input rate: ₹0.25 per million on DeepSeek V4.1 Flash, against ₹14 uncached. That is a 98% cut on the repeated part.
Sarvam also describes the engineering behind its numbers, per the AlphaSignal summary:
- A skipped clamping step in a GLM 5.3 Flash kernel pushed output deviations to 165%. A fused replacement kernel fixed it and ran about 40% faster.
- Moving GLM 5.3 from FP8 on SGLang to MXFP8 on vLLM raised uncached prefill from 3,700 to 7,600 tokens per second per GPU and halved p95 time to first token. Sarvam's thread puts it as "halves time to first token and doubles throughput."
- Aligning cache block sizes on DeepSeek V4.1 Flash 1017 lifted the cold-cache hit rate from 37.1% to 99.8%.
These are good signs of real operations work. They are still self-reported, and none of them tells you what your own prompts will see.
What does this mean for what you build or pay?
If you build in India, or serve Indian users, three things change.
- You can pay in rupees. Many Indian teams pay foreign inference vendors in dollars and carry exchange and card fees. A rupee invoice is simpler for finance teams.
- Latency can drop. A call that stays in India skips an overseas round trip. For voice and chat, the first-token time Sarvam quotes is what users feel. Test it from your own region.
- Data stays closer to home. Sarvam says requests run on Sarvam-operated GPUs in India, that zero data retention is available on request, that the managed service carries a 99.95% uptime SLA, and that regulated customers can deploy on-premises. Treat those as claims to confirm in a contract. The API docs page we read does not state residency terms.
Data crossing a boundary, showing why India-hosted LLM inference matters for data residency
Language cost matters too. Hindi and other Indian languages often need more tokens per sentence than English, as we showed in our tokenizer tax analysis. A lower per-token price helps most on those workloads. A model tuned for Indian languages may also need fewer tokens for the same text, but you should measure that on your own data.
How do you try it today?
Sarvam says you switch from OpenAI by changing the base URL and API key and setting the model field. A minimal call looks like this. It follows Sarvam's description of the API, and we have not run it ourselves.
from openai import OpenAI
client = OpenAI(
base_url="https://api.sarvam.ai/v2",
api_key="YOUR_SARVAM_API_KEY",
)
resp = client.chat.completions.create(
model="sarvam-105b",
messages=[{"role": "user", "content": "Explain UPI in two lines, in Hindi."}],
stream=True,
)
for chunk in resp:
print(chunk.choices[0].delta.content or "", end="")
The docs also show a cURL option that uses an api-subscription-key header, and a Sarvam SDK path. Token Factory uses the same API key and credits as the other Sarvam APIs. The login page for the console is gated, so you need an account first. Sarvam links migration guides for moving from OpenRouter, Fireworks AI and Together AI.
A short test plan, in order:
- Pick one real workload, such as a support-chat prompt or a coding-agent loop.
- Run 200 requests on your current provider and on
sarvam-105b, with the same prompts. - Record first-token time, tokens per second, cache-hit share and cost in rupees.
- Compare output quality on a held-out sample you can score.
- Only then move traffic.
What are the open questions?
- Independent benchmarks. Sarvam compares itself with Together, Fireworks, Baseten, DeepInfra and InferenceX. We have not seen outside replication.
- The 80% figure. It does not match the list prices we read. Ask Sarvam for the method.
- Introductory pricing. It can rise after 2026.
- Provisioned models. DeepSeek V4 Pro, Kimi K3.1 Pro and Qwen 3.8 27B need per-organisation enablement.
- Residency terms. Get the data-processing wording in writing before you send regulated data.
- Model quality. Cheaper tokens do not help if the model fails your task. Evaluate first.
For the model side of this story, read our coverage of Sarvam Prism, which is also served on Token Factory. For the bigger picture of Indian AI, see the Sarvam Epoch recap, the Sarvam models and API guide and the IndiaAI Mission status report. If you compare Asian open-weight models on price, our OpenRouter token share analysis gives the market context.
Related reading
- Sarvam Prism decision model
- Sarvam Epoch 2026 announcements
- Sarvam AI models and API guide
- India sovereign AI status
- Tokenizer tax for non-English languages
- Gemma 4 31B on Cerebras
- DeepSeek V4 Flash with Codex
- Asia AI models on OpenRouter
Official sources: Token Factory docs, Sarvam pricing and rate limits and Sarvam's launch thread on X.
Prices, limits and model lists reflect Sarvam's documentation and launch thread on November 16, 2026. Introductory pricing and model availability may change.
