AI & Agents

Sarvam Opens Token Factory, an India-Based Inference Platform

Leave a comment

Sarvam has opened Sarvam Token Factory to everyone, MarkTechPost reports. It is an inference platform, not a single model: Sarvam serves its own models and a set of open-weight models from NVIDIA B200 clusters in India, all behind one OpenAI-compatible API. The launch list includes Sarvam 105B (105B total, 9B active, 128K context), GLM 5.3 and GLM 5.3 Flash, DeepSeek V4.1 Flash 1017 and DeepSeek V4 Pro, Kimi K3 and Kimi K3.1 Pro, and Gemma 4 31B. Some of those run on shared pay-as-you-go endpoints. Others, including DeepSeek V4 Pro, Kimi K3.1 Pro, Qwen 3.8 27B, Clef, and EmbeddingGemma 2, need provisioned throughput. Enterprises can also ask about on-premise deployment. Sarvam says Sarvam 105B now fits on one B200 per replica.

The speed claims are Sarvam’s own. It reports the fastest first token from Mumbai in most of the cases it tested, and says DeepSeek V4.1 Flash serves more than 450 billion tokens a day with 95.7% of prompt tokens coming from cache. For a single user, it lists Sarvam 105B at 0.2 seconds to first token and 330 tokens per second. I have not verified those numbers, and I have not signed up for the API.

My take

An India-hosted, OpenAI-compatible endpoint is mainly interesting if your users or your compliance story want the traffic to stay there, or if the cache and routing work is as good as advertised. I would treat the leaderboard figures as a vendor’s launch post until I can run my own prompts against it.

God Bless, and Happy hacking!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.