Sarvam Launches India-Hosted Token Factory at 450 Billion Tokens a Day
Sarvam AI says Token Factory already serves 450 billion DeepSeek tokens daily from Indian GPUs, with 95.7% cache hits — though Together beats it on DeepSeek speed.
By Olivia Hart
4 min read

What's News
- Sarvam launched Token Factory on November 16, 2026, serving over 450 billion DeepSeek V4.1 Flash tokens per day.
- Introductory pricing: ₹10/₹25 per million input/output tokens for Sarvam 105B, with cached inputs at ₹0.25 per million.
- 95.7% of DeepSeek V4.1 Flash prompt tokens are served from cache; G3 cut peak GLM 5.3 prefill nodes from 58 to 34.
- In Sarvam's own test, Together returned DeepSeek first tokens in 0.71s and streamed 278 tokens/s, beating Sarvam Standard's 0.90s and 130 tokens/s.
- Enterprise deployments carry a 99.95% uptime SLA and zero-data-retention terms on India-based GPUs.
Sarvam AI says its new inference platform already serves more than 450 billion DeepSeek V4.1 Flash tokens per day. The Bengaluru company launched Sarvam Token Factory on November 16, 2026, offering its own and third-party open-weight models from GPU clusters in India through a single OpenAI-compatible API.
The launch lineup includes Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3.1 Pro and Gemma 4 31B. According to the API documentation, Token Factory reuses the same API key, SDK and credits as Sarvam's other APIs. Users switch models by changing the model field; OpenAI-compatible clients point to https://api.sarvam.ai/v2.
What does it cost?
Introductory pricing for Sarvam 105B runs ₹10 per million input tokens and ₹25 per million output tokens. The Standard endpoint for DeepSeek V4.1 Flash 1017 costs ₹14 and ₹55 respectively. Cached-input rates sit at ₹0.25 per million tokens for both. The catalogue states the introductory rates run through the end of 2026, and all rupee prices exclude GST.
Deployment options go beyond pay-per-token serverless endpoints. DeepSeek V4 Pro and Kimi K3.1 Pro are enabled per organization under Provisioned Throughput; the rest are serverless. Customers can pick shared or dedicated capacity. Sarvam also sells the inference optimization stack for on-premises deployment and says critical institutions already run it that way.
The company's product page claims inference runs on Sarvam-operated GPUs in India rather than being forwarded to the original developers' overseas APIs. Under a zero-data-retention arrangement, prompts and responses are not retained, and customer data does not train models for other customers. Enterprise deployments carry a contractual SLA of 99.95% uptime. The claim is contractual, not independently audited.
How does the cache architecture work?
The serving stack splits prefill, which reads the full prompt, from decode, which generates tokens one at a time, and runs each phase on separate GPU pools. The design keeps a new long coding prompt from stalling active decoding sessions, a known failure mode when both phases share a GPU. A cache-aware control plane routes each turn toward its existing context, and admission counts only uncached tokens.
A second cache tier, built on Moonshot's Mooncake-based G3 store, is accessible to every prefill server and retains prompts for hours or days depending on traffic. GPU memory alone holds a prompt for about four seconds, dropping to roughly two under load. That gap matters for agents that return after a tool call with nearly the same prompt.
Sarvam reports that 95.7% of DeepSeek V4.1 Flash prompt tokens are served from cache. G3 also cut peak GLM 5.3 prefill-node requirements from 58 nodes to 34. For Sarvam 105B, moving from BF16 weights to the 8-bit MXFP8 format plus speculative decoding raised output throughput roughly 2.2-fold and halved each replica from two B200 GPUs to one, with no accuracy loss, the company says. An orchestrator agent assigns optimization experiments to sub-agents, but an engineer approves every change before deployment.
How fast is it, really?
Sarvam's strongest numbers favor its smaller models. In a 72-hour median test from an AWS Mumbai client with a 10,000-token prompt, Gemma 4 31B returned its first token in 0.40 seconds and streamed at 308 tokens per second. The only other provider serving that model posted 2.00 seconds and 193 tokens per second, measured at its best tier for each metric.
A separate Sarvam test of Kimi K3.1 Pro on coding-agent traces using 16 B200 GPUs produced 6,177 tokens per second per GPU at roughly 70 tokens per second per user. That compares with 3,213 on InferenceX's published curve.
The DeepSeek comparison is less flattering. In Sarvam's own 10,000-token latency test, Together returned the first token in 0.71 seconds versus 0.90 seconds for Sarvam Standard. The company's price-versus-speed table lists Together at 278 output tokens per second against 130 for Sarvam Standard. Those figures were measured November 3–5, 2026, one request at a time, on each provider's fastest tier.
The takeaway: Token Factory makes the serving stack itself the product, pairing India-based processing and OpenAI-compatible migration with GPU savings from caching and per-model tuning. The published gains are real in selected workloads, but buyers should expect throughput and latency advantages to vary by model and endpoint.
Source: unite.ai
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,221 articles