Sarvam Prism

Architectural Overview & Capability Profile

Sarvam Prism is an open-weights foundation model developed by Sarvam AI specifically engineered for high-speed “System-1” decision making, policy evaluation, and agent triage. Rather than generating long-form conversational text token-by-token, Prism evaluates inputs against predefined categories, intent choices, and policy rules in a single forward pass, returning calibrated probabilities and classification rankings. On RankLLMs’s evaluation framework, Sarvam Prism achieves an overall rating of 48.2 / 100, standing out as one of the fastest decision models available.

Built as an optimized fine-tune of Sarvam 105B (a sparse Mixture-of-Experts architecture featuring 128 grouped experts), Prism operates over a 131,072-token context window under a permissive Apache 2.0 open-weights license.

Decision Benchmarks & Real-Time Latency

Evaluated across Sarvam’s comparative benchmark suite of 25 decision engines, Prism leads all models responding in under 100 milliseconds:

1. System-1 Decision Accuracy:
   • Average Decision Index Score: Achieves 71.8, leading high-speed open-weights classifiers.
   • Indic-Language Decision Benchmark: Scores 82.5, demonstrating native understanding and nuanced intent classification across Indian regional languages.
   • Jev Decision Index (37 benchmarks): Records 67.0.
2. Inference Latency:
   • Clocked at a median response time of just 44 milliseconds (p95 of 420 ms) when hosted on a single NVIDIA B200 GPU in FP8 precision.
   • Throughput reaches 210 decisions/tokens per second.

Deployment & Economics

Sarvam Prism is distributed via Hugging Face (sarvamai/prism) and available through Sarvam’s cloud API:

• Cloud API: $0.15 per 1M input tokens and $0.45 per 1M output tokens.
• Hardware Hosting: Deployable on a single 120GB+ VRAM accelerator (e.g. NVIDIA B200) for enterprise private VPC inference.