AI News

Trending story

Sarvam AI launches Sarvam Token Factory inference service

1 reports1 story updatesUpdated Nov 7 08:00

Understand this story

AI overview

On November 7, 2026, Sarvam AI launched the Sarvam Token Factory inference service. According to information published on the Sarvam AI Blog, the platform uses GPU clusters located in India and hosts Sarvam 105B as well as open-weight models such as GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro, and Gemma 4 31B, all invoked uniformly through an OpenAI-compatible API. On pricing, the standard endpoint for DeepSeek V4.1 Flash costs 14 rupees per million input tokens and 55 rupees per million output tokens; Gemma 4 31B takes 0.40 seconds to first token with a 10K-token prompt and outputs 308 tokens/s. The platform says it serves 450B+ tokens per day on DeepSeek V4.1 Flash, with 95.7% of prompt tokens hitting the cache. There are currently no subsequent developments.

AI-generated from public reports

Latest development · Report

Sarvam AI Launches Token Factory Inference Platform, Hosting Proprietary and Open-Source Models

Developments

1 confirmed developments

  1. Sarvam AI launches Sarvam Token Factory inference service

    Source:Sarvam AI Blog

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 7

  1. Sarvam AI BlogFirst-partySelected

    Sarvam AI Launches Token Factory Inference Platform, Hosting Proprietary and Open-Source Models

    Sarvam AI has released Token Factory, which uses GPU clusters located in India to host Sarvam 105B as well as open-weight models such as GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro, and Gemma 4 31B, all accessed through an OpenAI-compatible API. DeepSeek V4.1 Flash is priced at 14 rupees per million input tokens and 55 rupees per million output tokens on the standard endpoint; Gemma 4 31B takes 0.40 seconds to first token on a 10K-token prompt and outputs 308 tokens/s. The platform says it serves 450B+ tokens per day on DeepSeek V4.1 Flash, with 95.7% of prompt tokens hitting the cache.

Related stories

  1. Related storiesMistral releases Large 4 weights on Hugging Face
  2. Related storiesMiniMax releases the H3-Image image model and opens the weights
  3. Related storiesQwen Releases Qwen4-Max Preview Version
  4. Related storiesRecursive releases open-weight local coding model Flint