Sarvam Releases Prism, an Open-Weight AI Model That Picks an Answer in 44 Milliseconds
Sarvam AI released Sarvam Prism on November 3, 2026, an open-weight model built to make a fast, structured choice rather than write an answer word by word. Instead of generating text, Prism takes a situation, a question and a set of predefined options, then returns a probability for each option in a single forward pass. The company reports a median response time of 44 milliseconds on one NVIDIA B200 GPU and an average score of 71.8 across three decision benchmarks.
The release matters because much enterprise AI work is not open-ended writing. It is a chain of small decisions: route a support request, flag a document, check an agent run, or block a suspicious input. For those jobs, a model that returns a calibrated score for each option can be easier to automate than a chatbot-style answer that must then be parsed.
Sarvam describes Prism as a System One decision model, borrowing the term for fast, intuitive judgment. The company notes that TypeSafe created this category last month with its Jev model.
Prism is a LoRA fine-tune of Sarvam 105B, built on the Sarvam 105B-10B base. Each request contains a state, a question and any number of options that fit within the context window. A small scoring head assigns a probability to every option, with no token-by-token generation. Sarvam says the model can handle up to 131,072 tokens per question, so long inputs such as agent traces can be evaluated in one pass rather than split into chunks.
The speed claim rests on a series of engineering changes. Sarvam says Prism started at a median of 664 milliseconds per request in 16-bit precision across two B200 GPUs. Converting the weights to FP8 allowed the model to fit on a single GPU. The company also grouped all 128 experts into one kernel, captured the forward pass as a CUDA Graph, folded FP8 conversion into the expert kernels and merged the fine-tuning adapter into the model weights. Together, Sarvam reports, those steps cut the median to 44 milliseconds, about 15 times faster, with no drop in accuracy.
On accuracy, Sarvam tested Prism and 24 other decision models on three indices. The average score of 71.8 combines 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 says Prism has the highest average score among models answering in under 100 milliseconds in its 25-model test, although a few models score higher overall while taking between two and six times as long. These are Sarvam’s own measurements, and two API-only systems in the comparison appear with publicly reported timings rather than equivalent local measurements.
Calibration is a central part of the pitch: an enterprise can automate a decision only above a confidence threshold, and that is safe only if the confidence figure means what it says. For English answers given with at least 90 percent confidence, Prism expected 3.3 percent errors and made 2.2 percent, across 226,088 questions from 37 public benchmarks. For 20,108 questions in 10 Indian languages and Hinglish, it expected 5.4 percent errors and made 4.4 percent.
Prism also addresses a quieter consistency problem: changing the order of options can change some models’ answers. In Sarvam’s 2,000-question test, Prism changed 4.0 percent of answers when options were shuffled and scored as received. By placing options in a fixed internal order before scoring and returning each probability under its original label, Sarvam says, that figure fell to zero, with no extra computation and no accuracy loss in the evaluation.
The company tested Prism on agent oversight and security checks as well. On 700 recorded agent runs covering 1,770 questions, Prism scored 78.5 percent overall, including 75.0 percent for judging whether a task succeeded, 86.4 percent for deciding whether human review was needed and 69.6 percent for identifying the agent responsible for a failure. Sarvam says a full agent run takes about half a second on a single GPU, compared with a median 2.5 seconds per run for Claude Haiku 5.5 called through its API. On an internal cybersecurity benchmark of 5,517 questions across nine tasks, Prism scored 89.0 percent, ahead of Clef at 84.5 percent, Jev 1.14 at 78.6 percent and Microsoft-Decision-1 at 76.1 percent. Sarvam plans to release that evaluation set publicly in the coming weeks.
For deployment, Prism’s weights are available on Hugging Face under the Apache 2.0 licence, and hosted access is available through the Sarvam API on Sarvam Indus. Sarvam says enterprises can train the model on their own data and run it on their own infrastructure. A local deployment runs in FP8 on one GPU with at least 120 GB of memory. Reported enterprise uses include customer-support routing, security checks and evaluation of AI-agent activity.
Why this matters beyond one benchmark table: if decision models deliver calibrated, order-stable scores at millisecond speed, they could become quiet infrastructure inside agents, support systems and security filters, rather than the visible chatbot on top. The open weights make that proposition testable on private workflows. The limits matter too: the headline scores come from Sarvam’s own testing, the security benchmark is internal and not yet public, and some rival comparisons mix measured and reported timings. Prism should be treated as a promising decision layer to evaluate on real tasks, not as a settled verdict on which decision model is best.
