Sarvam Token Factory
Intro
Sarvam Token Factory is the inference service that the Indian AI company Sarvam opened to everyone on 16 November 2026. The official account @SarvamAI posted at 2026-11-16T13:37:16Z (21:37:16 on 16 November 2026, Asia/Shanghai): “Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability.” When this scan re-checked the original post it had about 2,673 likes, 316 reposts, about 595,000 impressions and 503 bookmarks. That introducing post sat inside the 00:32 Asia/Shanghai window on 17 November, but was not stored then; this slot backfills it after checking the post and the official blog.
The same-day official blog post Introducing Sarvam Token Factory describes Token Factory as Sarvam’s inference stack for serving its own models and leading open-weight models from GPU clusters located in India, and says that on most of the models it serves, it matches or beats leading global inference providers on both speed and cost. The catalogue listed in the blog 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, all through one OpenAI-compatible API. The company also writes that it runs India’s largest B200 cluster, already runs the same inference-optimisation stack on-premise for critical institutions, and offers that deployment option to enterprises.
Follow-up posts in the same official thread add two details: DeepSeek V4.1 Flash 1017 is served on an Interactive endpoint (about 200 tokens per second per user, for coding agents) and a Standard endpoint (for background and batch work, with introductory pricing comparable to GPT-6 Luna and DeepSeek’s own API); and Gemma 4 31B is described as an efficient real-time conversation model that combines agentic capability with Indian-language support, with Sarvam claiming a 5× sooner first token and 1.6× more tokens per second than the only other provider serving it. The “Start building” link in the official posts points at Token Factory on Indus (indus.sarvam.ai). Do not confuse this with the already stored term Sarvam Prism, an earlier Sarvam model/product line.
Advertisement
Explain
Token Factory is an inference platform, not a new foundation-model name. The blog summarises it as follows. On cost, Sarvam says it offers globally competitive pricing on every model it serves. On performance, it says Sarvam 105B brings together its inference optimisations for conversation at the lowest cost, returning a first token in 0.2 seconds with single-user throughput of 330 tokens per second. On latency, Gemma 4 31B is reported to return its first token in 0.40 seconds on a 10K-token prompt measured from Mumbai, which the blog says is 5× faster than the only other provider serving it (2.00 seconds), with output speed of 308 tokens/s versus 193. On throughput, Kimi K3.1 Pro is said to deliver up to 1.9× tokens per GPU at the same per-user speed as InferenceX’s published results, measured on coding tasks. These figures all come from the official blog. The Gemma comparison is described as a 72-hour median from a client in AWS Mumbai, with the other provider’s bar being the best of its tiers on that metric.
On pricing, the blog gives introductory rates for DeepSeek V4.1 Flash: Standard at ₹14 per million input tokens and ₹55 per million output tokens; Interactive at ₹25 and ₹105. Both endpoints serve the model at native precision. The company also describes a three-tier cache with cache-aware routing and admission for newer attention architectures, and says that on DeepSeek V4.1 Flash it serves 450 billion-plus tokens a day, with 95.7% of prompt tokens served from cache. The second half of the blog goes into how they run experiments across the stack with profilers, a simulator calibrated on their servers and an evaluation suite, with an engineer approving every change before deployment. It also credits upstream work from teams behind vLLM, SGLang, Dynamo, Mooncake, NIXL, DeepEP, DeepGEMM and FlashInfer.
For enterprises, the official message is more models to choose from, and an AI stack they can own without lock-in. Besides the Token Factory cloud service, the same optimisation stack is already run on-premise for critical institutions and is offered as an enterprise option. Data residency in India and rupee pricing are the differences that official materials and news coverage keep returning to.
Usage
According to the official blog, Sarvam Token Factory is available today to developers and enterprises. You call the models in the catalogue through one OpenAI-compatible API, and Sarvam says it will keep adding new open-weight models from day one and expanding compute as demand grows. The starting point in the official introducing thread is the “Start building” link, which leads to Token Factory on Indus (indus.sarvam.ai). The fullest product description is the official blog post.
The blog is specific about endpoint choice. For coding agents that want higher per-user speed, the Interactive endpoint of DeepSeek V4.1 Flash is officially described as about 200 tokens per second per user. For background or batch work, Standard is the lower introductory price tier. For real-time conversation in Indian languages, the official recommendation is Gemma 4 31B, with the Mumbai first-token and throughput comparison quoted above. Sarvam 105B is presented as the company’s own combination of conversation speed and cost. If a cloud endpoint is not enough, the same inference-optimisation stack is offered for on-premise deployment.
Because the API is OpenAI-compatible, apps, agents and gateways that already speak the OpenAI interface can in principle switch by changing the base URL, model name and key. The live catalogue, quotas and authentication are whatever the official site lists at the time; this page does not invent endpoint paths or a sign-up sequence that the official materials did not spell out. Sarvam also says it is building more tools for developers to build and deploy agents on Token Factory.