Sarvam AI has released Sarvam Prism, an open-weight decision model built on its 105-billion-parameter Sarvam 105B foundation model, with weights under the Apache 2.0 licence. Sarvam says Prism responds in a median of 44 milliseconds on an NVIDIA B200 GPU, according to ANI and CXOToday reports dated 4 November 2026.
Bengaluru-based Sarvam AI describes Prism as a model for “fast, structured decisions”. Rather than generating text to reach an answer, it scores the available options in a single pass and returns a probability for each, Inc42 reported on 5 November 2026.
What is Sarvam Prism and what is it for?
Prism is aimed at enterprise workflows made of many small decisions. Use cases named in the reports include customer-support routing, document intelligence, security checks, prompt-injection detection and evaluation of AI-agent activity. Sarvam said in its release: “We are releasing Prism with open weights so that enterprises can train this calibrated base on their own data.”
How does Sarvam Prism perform on benchmarks?
All figures below are reported by Sarvam and were not independently verified in the articles.
- An average score of 71.8 across three evaluation indices, the highest among the 25 decision models Sarvam tested that respond in under 100 milliseconds.
- 82.5 on Sarvam’s internal Indian-language benchmark, versus 75 for Perplexity Decider v1.2 and 70.8 for Jev 1.14.
- 89% accuracy on an internal cybersecurity benchmark of 5,517 questions, versus 84.5% for Clef and 78.6% for Jev 1.14.
- For answers given with at least 90% confidence, error rates of 2.2% in English and 4.4% in Indian languages, both below the predicted error rates.
- Unchanged answers across all 2,000 questions when answer options were arranged in a fixed order before scoring.
What Sarvam Prism means for Indian businesses
Because the weights are open under Apache 2.0, enterprises can adapt the model and run it on their own infrastructure. It is also available through Sarvam’s API. Inc42 reported on 4 November that the weights are on Hugging Face and that local deployment requires a GPU with at least 120 GB of memory, a point that matters for firms weighing self-hosting against the API.
Industry Reaction and Expert Commentary
Early independent views are mixed. Sandipan Kundu, builder of Tensorant, said on X that the model appeared capable but questioned the 44 ms median latency claim, citing a p95 latency of 420 ms in his testing. He also flagged the cost of self-hosting a model of this size.
Faizan Shaik of Atlan reported that on Atlan’s Decision Bench, Prism scored 91.5% across 1,071 decisions on 36 public datasets, placing 12th among 22 models, with a median latency of 65 ms on a B200 GPU. Ananthanarayanan S, MD India at Articul8, said Sarvam’s launch adds to the range of general-purpose models, while Articul8’s Bivia targets a different enterprise need.
Prism enters a growing field of decision models. Inc42 names Laya from ConvAI Innovations, Sieve from PurpleTalk’s Sthānika AI, Bivia from Articul8 and Jev from TypeSafe AI, and notes that Microsoft and Perplexity have also introduced decision-model offerings.
What Happens Next?
Inc42 notes that Sarvam’s results are company-reported and that independent testing across comparable workloads is needed to show how they hold up in production. The articles do not give pricing for Sarvam’s API.
Frequently Asked Questions
What is Sarvam Prism?
It is an open-weight decision model from Sarvam AI, built on the 105B-parameter Sarvam 105B model, that scores a fixed set of options and returns a probability for each.
Is Sarvam Prism open source?
Its weights are released under the Apache 2.0 licence, so developers can adapt it and run it on their own infrastructure. It is also available through Sarvam’s API.
How fast is Sarvam Prism?
Sarvam reports a median response time of 44 milliseconds on an NVIDIA B200 GPU. Atlan measured a median of 65 ms on a B200, and Tensorant’s Sandipan Kundu cited a p95 latency of 420 ms in his testing.
Source: ANI via Devdiscourse, 4 November 2026; CXOToday, 4 November 2026; Inc42, 4 and 5 November 2026.
Leave a comment