Sarvam Prism

// Overview

Sarvam Prism is a decision model from Sarvam AI, announced on the Sarvam blog on November 3, 2026, and built on Sarvam 105B, the company's Apache-2.0 open-weight mixture-of-experts model. Sarvam calls it a "System One" model. It does not write text. A request carries a state (text, or a JSON object or array) plus typed questions, and Prism returns a probability for every option in a single forward pass. The question types are yes/no, choice (keyed options) and score (ordered levels). The weights are on Hugging Face under Apache 2.0, and the model is also served on the Sarvam API.

Sarvam scored Prism against 24 other decision models on three indices: the Jev Decision Index (37 public benchmarks, 226,088 questions), a recreation of the Artificial Analysis Decision Index, and an internal Indian-language set (the same benchmarks in 10 Indian languages and Hinglish, 20,108 questions). Prism averages 71.8 (67.0, 65.9 and 82.5 respectively) with a 44 ms median response on one B200 GPU. That puts it behind mammoth-1 (75.2) and GLiDE 1.2 (73.2) on the average but makes it, in Sarvam's words, "the highest-scoring model under 100 ms across the 25 models we tested." On an internal benchmark of 5,517 cybersecurity questions (phishing, jailbreak detection, MITRE attack types) it scores 89.0%, ahead of Clef (84.5%), Jev 1.14 (78.6%) and Microsoft-Decision-1 (76.1%). Judging 700 recorded agent runs, it scores 78.5%.

The release post describes the recipe. Prism is a LoRA fine-tune of the "Sarvam 105B-10B base", trained with two equally weighted losses: cross-entropy with 5% label smoothing and a Brier loss. Sarvam credits those two terms for its calibration. At 90% or higher confidence, Prism expected 3.3% errors in English and made 2.2%, and expected 5.4% in Indian languages and made 4.4%. Across 2,000 questions run in four shuffled option orders, its answers stayed the same. Inference began at 664 ms per median request in 16-bit precision on two B200 GPUs. FP8 weights, one grouped kernel for all 128 experts, CUDA Graphs and a merged adapter brought that to 44 ms on one GPU.

To self-host, the Hugging Face repository sarvamai/prism needs one NVIDIA GPU with FP8 support and about 120 GB of free memory (tested on a B200), plus about 106 GB of disk. It ships its own loader (prism.py with Prism.from_pretrained and model.decide) because the custom FP8 layout does not load in AutoModelForCausalLM or general-purpose inference servers. Self-hosted, a question can run to 131,072 tokens. On the Sarvam API, the model id sarvam-prism is called at POST https://api.sarvam.ai/v2/systemone, with a limit of 32,768 tokens per question, 2 to 255 options per choice question and 2 to 10 levels per score question. The price is ₹4 per million input tokens and nothing for output.

// Benchmarks

This model's scores

1. Average of three decision indices (Sarvam)71.8
2. Jev Decision Index67
3. Artificial Analysis Decision Index (Sarvam recreation)65.9
4. Indian-language decision index (Sarvam)82.5
5. Cybersecurity benchmark, 5,517 questions (Sarvam internal)89%
6. Agent trace judging, 700 runs (Sarvam)78.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Sarvam API, the same on every plan

Pricing source ↗

// Strengths

• 71.8 averaged over the Jev Decision Index, an Artificial Analysis Decision Index recreation and an Indian-language set
• 44 ms median response on one B200 — the highest-scoring of 25 tested decision models under 100 ms
• 82.5 on the Indian-language index: English, 10 Indian languages and Hinglish
• 89.0% on Sarvam's 5,517-question cybersecurity benchmark, ahead of Clef, Jev 1.14 and Microsoft-Decision-1
• Calibrated confidence: 2.2% actual errors against 3.3% expected at ≥90% confidence in English
• Apache-2.0 weights on Hugging Face, plus a hosted API at ₹4 per 1M input tokens with free output

// Best for

• Reach for Prism to route, triage or classify requests in English and Indian languages, where one call returns a probability for every option you can threshold.
• Reach for it as a security guardrail: phishing and jailbreak detection, or mapping an artifact to MITRE attack types.
• Reach for it to judge agent traces — task success, whether a run needs human review, and failure attribution — with up to 131,072 tokens per question when self-hosted.
• Pick mammoth-1 or GLiDE 1.2 when the last points of average accuracy matter more than latency.
• Look elsewhere when you need generated text, image input, or a model that fits on a GPU with less than about 120 GB of memory.

// How to access