Sarvam Token Factory Opens India-Hosted AI Inference for Sarvam 105B, DeepSeek, Kimi and Gemma

Sarvam AI has opened Token Factory to everyone, serving Sarvam 105B plus open-weight models such as DeepSeek V4.1 Flash, Kimi K3.1 Pro, GLM 5.3 and Gemma 4 31B from GPU clusters in India through one OpenAI-compatible API.

Christopher Ort
By Christopher Ort · 2 min read
Share
A data center building at dusk with cooling units on the roof.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. How it works
  6. Why it matters
  7. What's still open
  8. Sources
  9. i10x Perspective

Indian AI company Sarvam has opened Sarvam Token Factory to everyone, an inference service that serves its own Sarvam 105B model and leading open-weight models from GPU clusters in India. The launch was announced on 16 November 2026 and offers all models through one OpenAI-compatible API.

i10X Quick Take

Quick Take

  1. Models: 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.
  2. Location: served from GPU clusters in India, which Sarvam says include the country’s largest B200 cluster.
  3. Price example: DeepSeek V4.1 Flash starts at introductory rates of ₹14 per million input tokens and ₹55 per million output tokens on the Standard endpoint.
  4. Speed claim: Sarvam 105B returns its first token in 0.2 seconds at 330 tokens per second for a single user.
  5. API: OpenAI-compatible, using an existing Sarvam API key.

In plain English

Developers usually call AI models through providers in the US or Europe. Sarvam now runs popular open models from China, the US and its own lab on servers in India, at prices it says match or beat global providers. For Indian companies, that means data stays in the country and responses arrive faster.

What happened

Sarvam wrote in its launch post: “On most of the models we serve, we match or beat the leading inference providers globally on both speed and cost.” The company also runs the same inference stack on-premises for critical institutions.

Background

Open-weight models such as DeepSeek V4.1 Flash and Moonshot’s Kimi have become a large share of real-world AI usage, and inference providers compete hard on price and latency. Sarvam is one of India’s best-known AI labs and has trained its own large models.

How it works

  • Two endpoints for DeepSeek V4.1 Flash: Interactive, at about 200 tokens per second per user for coding agents, priced at ₹25 input and ₹105 output per million tokens, and Standard for background and batch work.
  • Latency: Sarvam says Gemma 4 31B returns its first token in 0.40 seconds on a 10K-token prompt, 5 times faster than the only other provider serving it.
  • Throughput: Kimi K3.1 Pro delivers up to 1.9 times the tokens per GPU at the same per-user speed as InferenceX’s published results, on coding tasks.
  • Caching: a three-tier cache with cache-aware routing. On DeepSeek V4.1 Flash, Sarvam serves more than 450 billion tokens a day, with 95.7% of prompt tokens served from cache.
  • Method: measurements come from a client in AWS Mumbai, median over 72 hours, using a recreation of Artificial Analysis’s method.

Why it matters

For Indian developers and companies, local hosting can help with data residency rules and lower latency.

For global inference providers, a regional player is entering a price war on open models.

For model makers, more hosts in more countries widen the reach of open weights.

What’s still open

  • Independent checks of the speed and cost claims, which come from Sarvam’s own tests.
  • Prices outside the introductory period.
  • Service levels and quotas for heavy enterprise use.

Sources

  • Sarvam AI, “Introducing Sarvam Token Factory,” blog post, 16 November 2026
  • Sarvam API documentation, Token Factory and open-weight models
  • The Economic Times and India Today, reports on the launch, 17 November 2026

i10x Perspective

Sovereign AI is often discussed as training national models. Sarvam shows the other half: serving the world’s best open models at home, fast and cheap. More on our DeepSeek company page and in the Infrastructure section.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    DeepSeek-V4.1-Flash-1017

    1M context · open weights

    api-docs.deepseek.com (source, opens in a new tab)
  2. Release

    DeepSeek-V4.1-Flash

    1M context · $0.3 / $1.2 per 1M tokens · Superseded 17 Oct 2026 by V4.1-Flash-1017

    api-docs.deepseek.com (source, opens in a new tab)
  3. Release

    DeepSeek-V4-Pro (GA)

    1M context · $1.32 / $3.96 per 1M tokens

    api-docs.deepseek.com (source, opens in a new tab)
  4. Release

    DeepSeek-V4-Flash

    Retired Oct 2026 (served by V4.1-Flash-1017)

    api-docs.deepseek.com (source, opens in a new tab)
  5. Funding

    DeepSeek closes first external round of over $7B

    First external round · ¥50B+ at $50B valuation

    Per The Information via Reuters: >¥50B (~$7.4B) at a post-money valuation above $50B, with an unusual deal structure.

    reuters.com (source, opens in a new tab)
Full DeepSeek timeline ›

Editor's picks

  1. Remanence Raises $30 Million Seed to Turn Companies’ Agent Data Into AI Models They Own
  2. Astra for Law: OpenAI Opens Its Legal AI to All ChatGPT Enterprise Customers
  3. Proton Lumo Desktop Brings a Privacy-First AI Assistant to Windows, Mac and Linux
  4. GitHub Copilot Evals Lets Developers Test Their Own Agents as Code Review Adds Adaptive Effort and Secret Detection
  5. Copilotforce: Microsoft and Salesforce Bring CRM Data and 38 Sales Skills Into Copilot