Tool Suggestions
Sarvam Prism: An Open-Weight Decision Model That Answers in 44 Milliseconds
Sarvam AI released Prism on November 3, a System One decision model built on Sarvam 105B. It does not generate text. You give it a state, a question and a list of options, and it returns a probability for each option in a single forward pass, in a median 44 milliseconds on one NVIDIA B200 GPU. The weights are open under the Apache 2.0 licence, so enterprises can fine-tune the model on their own outcomes and run it on their own infrastructure.
The category is new enough to need explaining. Decision models are the "System One" idea from Kahneman's thinking-fast-and-slow distinction: fast, intuitive judgement over a fixed set of choices, rather than slow, deliberative reasoning. TypeSafe created the category last month and named its Jev model after it. Prism is Sarvam's entry: a LoRA fine-tune of Sarvam 105B, the company's 105B-10B mixture-of-experts base, trained in India under the IndiaAI Mission and announced at the India AI Impact Summit 2026.
How it works
A normal language model answers a multiple-choice question by generating tokens one at a time, which is slow and produces text you then have to parse. Prism instead reads the state, the question and the candidate options within a 131,072-token window in one forward pass, and outputs calibrated probabilities through a lightweight scoring head. A split attention mask reads the state and question bidirectionally while isolating each option until the final answer token, which Sarvam says adds roughly 3 points on the Jev Decision Index and runs attention 2.6 times faster above 2,048 tokens.
The speed number comes from engineering, not just the architecture. Converting the weights to FP8, at 106 GB on disk and 120 GB of video memory, grouping 128 experts into one kernel, and capturing a CUDA graph cut median latency from 664 milliseconds on two 16-bit B200 GPUs to 44 milliseconds on one B200, a 15.1-times improvement. The 95th-percentile latency is 420 milliseconds, which is the number that matters if you are putting this in front of users and care about the slow tail, not just the median.
The numbers, with the caveat attached
Sarvam reports an average score of 71.8 across three decision indices: 67.0 on Jev, 65.9 on an Artificial Analysis recreation, and 82.5 on Indian languages. That is, by Sarvam's account, the highest average score among decision models that respond in under 100 milliseconds, across the 25 decision models it tested, ahead of Perplexity Decider v1.2 at 68.0 in 117 milliseconds and Jev 1.14 at 64.5 in 196 milliseconds.
Label that plainly: every benchmark figure and comparison in the announcement comes from Sarvam's own testing. There is no independent evaluation yet, and the Jev Decision Index itself is a new, vendor-associated benchmark. The honest way to read the claim is that Prism sets a new point on the speed-accuracy frontier among decision models Sarvam tested, and the open weights let anyone verify the accuracy half on their own data.
What it is for, and how to get it
The intended uses are enterprise workflows where decisions are fast, structured and repeated: customer support routing, document analysis, classification, moderation, triage, security checks, and evaluation of AI agents. The point of a decision model is that a probability is directly usable in a workflow — route to the highest-probability queue, escalate below a confidence threshold, log the distribution — where generated text would need parsing and would arrive hundreds of milliseconds later.
Prism's weights are on Hugging Face under Apache 2.0, and it is also available as an API on Sarvam Indus, though per-request API pricing was not printed in the announcement. If you already read DigestX's coverage of Nace Drex Max, note the difference: Drex Max is a dense 27B decision model that runs on a Mac, while Prism is a 105B mixture-of-experts LoRA aimed at server-side speed, at 44 milliseconds on a B200. Same category, different trade: Drex Max optimises for running anywhere, Prism optimises for answering fast at scale.