Sarvam AI Releases Open-Weight Prism Decision Model

Sarvam AI released Sarvam Prism on November 3, 2026, an open-weight model for enterprise decisions that uses Sarvam 105B as its base and scores predefined options without token-by-token generation. The company reports a 44-millisecond median response time on one NVIDIA B200 GPU and the highest average score among models answering in under 100 milliseconds in its 25-model evaluation.

The weights are available on Hugging Face under the Apache 2.0 licence, with hosted access through the Sarvam API on Sarvam Indus. A deployment requires one GPU with at least 120 GB of memory, with Prism running in FP8. It can process up to 131,072 tokens per question, allowing long inputs such as agent traces to be evaluated in one pass.

A Single-Pass Decision Model

Each request contains a state, a question and any number of options that fit within the context window. Prism processes that input in one forward pass, and a small scoring head assigns a probability to every option. Sarvam created the model by applying LoRA fine-tuning to its 105B system, adapting it for enterprise decisions rather than free-form text generation. Its tested uses include customer-support intent routing, model routing, document intelligence and checks on agent runs.

Two equally weighted training objectives shape those probabilities. One uses cross-entropy with 5% label smoothing to discourage complete certainty; the other uses Brier loss, which measures squared probability error and makes confident mistakes costly. Sarvam attributes Prism’s calibration to that combination. A custom attention mask also keeps options isolated from one another during scoring. The state and question are read bidirectionally, while only the answer token sees and compares all the options.

Accuracy and Latency Measurements

The reported average combines three indices: the Jev Decision Index, covering 37 public benchmarks; Sarvam’s recreation of the Artificial Analysis Decision Index, which substitutes curated difficult questions with an “I don’t know” option for the original private knowledge test; and an internal Indian-language benchmark. Response times were medians over 1,000 requests sent sequentially, using each model developer’s own inference code where Sarvam could run the model on its B200 hardware.

Prism scored 71.8 on average in Sarvam’s testing, with component results of 67.0 on Jev, 65.9 on the Artificial Analysis recreation and 82.5 on Indian languages. The higher-scoring mammoth-1 achieved 75.2, but its median response was 284 milliseconds. Two API-only systems, Jev 1.14 and Microsoft-Decision-1, appear in the comparison with publicly reported timings rather than equivalent local measurements, so the chart does not use one measurement method for every model.

The starting configuration ran at a median 664 milliseconds in 16-bit precision across two B200 GPUs. Converting the weights to FP8 allowed Prism to fit on one GPU. Sarvam also grouped all 128 experts into one kernel, captured the forward pass as a CUDA Graph, fused FP8 conversion into the expert kernels and merged the fine-tuning adapter into the model weights. Together, these changes produced the released single-GPU latency, and Sarvam reports no accuracy loss.

Calibration and Agent Evaluation

The calibration tests address whether a workflow can rely on confidence thresholds when automating decisions. At confidence of at least 90% in English, Sarvam reports that Prism expected 3.3% errors and made 2.2%, across 226,088 questions from 37 public benchmarks. For 20,108 questions in 10 Indian languages and Hinglish, it expected 5.4% errors and made 4.4%. Actual errors stayed below predicted errors at every tested threshold from 50% to 95% in both language groups.

Option order presents a separate consistency problem. In Sarvam’s 2,000-question test, Prism changed 4.0% of answers when options were shuffled into four orders and scored as received. Placing the options in a fixed internal order before scoring, then returning each probability under its original label, eliminated those changes. That step added no computation and caused no accuracy loss in the evaluation.

For agent oversight, Sarvam evaluated five models on the same 700 recorded runs and 1,770 questions. Prism’s reported overall accuracy was 78.5%, including 75.0% for judging task success, 86.4% for deciding whether human review was needed and 69.6% for identifying the agent responsible for a failure. The answers were checked against expert annotations or automatic environment outcomes, never another AI model’s judgement. Prism reads a trace once and scores five evaluation questions together, taking about half a second per run on one GPU.

Sarvam reports 89.0% accuracy for Prism on an internal cybersecurity benchmark of 5,517 questions across nine tasks and plans to expand and publicly release that evaluation set. Clef scored 84.5%, Jev 1.14 scored 78.6% and Microsoft-Decision-1 scored 76.1% in the same results. Built from vulnerabilities published since July 2026, the tasks include identifying phishing campaigns and jailbreak attempts and classifying attack types within MITRE frameworks.

Prism’s central proposition is a decision layer that enterprises can train and run on their own infrastructure, with confidence estimates designed for automated thresholds. The open weights make it possible to test that proposition on private workflows. The planned release of the security evaluation would also move one of the announcement’s internal tests toward public scrutiny.