Introducing Sarvam Token Factory
Introducing Sarvam Token Factory
Sarvam launched Token Factory, an inference stack running from GPU clusters in India that serves Sarvam 105B and open-weight models including GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Gemma 4 31B through one OpenAI-compatible API. The company reports matching or beating leading global inference providers on speed and cost for most models, citing figures such as Gemma 4 31B returning its first token in 0.40 seconds on a 10K-token prompt and 95.7% of DeepSeek V4.1 Flash prompt tokens served from cache. It also details engineering work across the inference engine, a three-tier cache and a cache-aware control plane, including moving GLM 5.3 to MXFP8 on vLLM, which doubled uncached prefill per GPU.
The post pairs vendor performance claims with unusually detailed engineering notes on caching, kernel fixes and routing, useful for readers evaluating inference providers or building agent workloads.
Source: Sarvam AI Blog · sarvam.ai