Sarvam Prism

Sarvam Prism is Sarvam AI's System One decision model, released November 3, 2026. It is a LoRA fine-tune of Sarvam 105B that scores fixed options instead of generating prose: you send a state, a question, and candidate answers, and it returns calibrated probabilities for yes/no, choice, and ordered-score questions in a single forward pass.

Across 25 decision models Sarvam timed, Prism posts a 71.8 average score with a 44 ms median on one B200—the best average among models under 100 ms—and about 82.5 on their Indian-language index. Weights ship Apache-2.0 on Hugging Face as sarvamai/prism, with a hosted Indus API as well.

Problem

Enterprise workflows are full of routing, triage, escalation, and safety checks. Chat LLMs that write a paragraph you then parse are slow and hard to calibrate. Prism isolates options in the attention mask, compares them at an answer token, and trains with cross-entropy plus Brier loss so confidence tracks error rates.

Versus peers

Comparisons include Jev, Clef, Microsoft-Decision-1, Perplexity Decider, mammoth-1, and others. Prism markets the accuracy–latency Pareto under 100 ms, plus conservative calibration and order-invariant answers (options are scored in a fixed key order).

Limits

• Needs ~120 GB FP8 GPU memory (B200/H200 class).
• Custom weight layout—not a drop-in for generic HF causal LM loaders.
• A decision head, not a general chatbot.

Who it is for

• Teams automating intent routing, agent-trace judging, jailbreak/guardrail checks, or model routing with explicit confidence thresholds.
• Products that need strong English + Indian-language / Hinglish decisioning.

Getting started

1. Read the launch post and Decisions API docs.
2. Download sarvamai/prism and run the bundled example.py on one FP8 GPU.
3. Or call the hosted API with a state plus questions map and gate automation on returned probabilities.