Sarvam AI releases Prism, an open-weight decision model with a 44 ms median response

Sarvam AI released Sarvam Prism on 3 November 2026, an open-weight decision model that scores a fixed list of options instead of writing text. The Bengaluru company said in a blog post that the model has a median response time of 44 milliseconds on one NVIDIA B200 GPU and scored 71.8 on average across three indices.

A decision model, not a chatbot

Sarvam says Prism scores a fixed list of options and returns probabilities.

Sarvam says each request carries a state, a question and any number of options. Prism handles that input in a single forward pass. A small scoring head assigns a probability to every option, and a custom attention mask keeps options isolated. Sarvam made the model by applying LoRA fine-tuning to Sarvam 105B.

Sarvam also lists model routing as a tested use.

Accuracy and speed

Sarvam Prism scored 71.8 on average across three indices: 67.0 on the Jev Decision Index, 65.9 on Sarvam's recreation of the Artificial Analysis Decision Index and 82.5 on an internal Indian-language benchmark.

Sarvam said that was the highest average among the 25 decision models in its evaluation that answered in under 100 ms, a set that includes Prism and 24 competitors. It ran the tests on the same type of chip, though two API-only systems, Jev 1.14 and Microsoft-Decision-1, used publicly reported timings rather than local measurements.

It is not the most accurate model in the table. Mammoth-1 scored 75.2 but had a 284 ms median response. Sarvam said its starting configuration ran at a median 664 ms in 16-bit precision across two B200 GPUs; converting the weights to FP8 let the model fit on one GPU.

Calibration and language tests

At confidence of at least 90% in English, Sarvam expected 3.3% errors. It 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% error and made 4.4%. In a separate 2,000-question test, Prism changed 4.0% of answers when options were shuffled into four orders and scored as received. Sarvam said placing options in a fixed internal order before scoring, then returning each probability under its original label, eliminated those changes.

On Sarvam's internal Indian-language benchmark, Sarvam Prism scored 82.5. Perplexity Decider v1.2 scored 75 and Jev 1.14 scored 70.8. On an internal cybersecurity benchmark of 5,517 questions across nine tasks, it recorded 89% accuracy. Clef scored 84.5% and Jev 1.14 scored 78.6%.

Open weights, one GPU

Sarvam released the weights on Hugging Face under Apache 2.0, with hosted access through the Sarvam API on Sarvam Indus. The company said deployment needs one GPU with at least 120 GB of memory.

Sarvam said Prism handles up to 131,072 tokens per question. That is enough to evaluate a long agent trace in one pass. Sarvam said enterprises can train the model on their own data. It named customer support routing, document intelligence, security checks and evaluation of AI-agent performance as early applications.

What is not settled

Sarvam ran the comparison itself. The average blends three indices: the Jev Decision Index, Sarvam's recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark.