Sarvam Token Factory Opens India-Hosted Inference to Eight Models
Sarvam has made its Token Factory inference platform publicly available. The service serves both Sarvam's proprietary models and leading open-weight models from GPU clusters located in India, giving developers a single OpenAI-compatible API for every model. For India-based teams, this means low-latency token generation without routing traffic through overseas clouds.
TL;DR
- Scale: Eight models at launch, including Sarvam 105B (105B total parameters, 9B active, 128K context), GLM 5.3, DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro, and Gemma 4 31B.
- Infrastructure: NVIDIA B200 clusters in India. Sarvam 105B now runs on a single B200 per replica. On-premise deployment is available for enterprises.
- Performance: The fastest time-to-first-token from Mumbai in most tested cases, according to Sarvam's internal measurements.
- Standout result: Kimi K3.1 Pro reaches 6,177 tokens/s per GPU versus 3,213 for InferenceX—a 1.9x advantage—at 70 tokens/s per user.
What Is Sarvam Token Factory?
Sarvam Token Factory is a managed inference stack rather than a single model. Sarvam first previewed the service as Sarvam Inference at its Epoch event. The launch includes eight models: Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3, Kimi K3.1 Pro, and Gemma 4 31B.
The documentation organizes access to these models through a unified, OpenAI-compatible endpoint, allowing teams to switch between or route across models without rebuilding integrations. Deployment options and specifications are available through GPULeaderboard.com's Chrome extension.
Why India-Hosted Inference Matters in 2026
As data-residency requirements tighten and latency-sensitive applications proliferate across enterprise AI and agentic workflows, sovereign inference infrastructure has moved from a compliance checkbox to a performance imperative. Hosting inference domestically lets Indian organizations keep data within the country while serving real-time assistants, voice agents, and multilingual applications at lower round-trip times.
Sarvam's choice of NVIDIA B200 hardware and single-GPU-per-replica efficiency for its 105B flagship reflects a broader 2026 trend: inference providers are competing less on raw model size and more on throughput per GPU, tokens per user, and cost efficiency at production scale. The move also positions Sarvam to capture enterprise demand for on-premise deployments where organizations retain full control over both data and infrastructure.
via MarkTechPost
