Sarvam AI releases Prism decision model

AI · First reported 3 Nov, 14:14 · Updated 15:16

Image: Wikimedia Commons · NASA Johnson Space Center / Jonny Kim · Public domain

Sarvam says Prism, a decision-focused model built on Sarvam 105B, returns probabilities for fixed answer options in a single pass. The company reports a 44-millisecond response time and an average score of 71.8 across three benchmarks; the model’s weights are available under the Apache 2.0 licence.

Prism is a LoRA fine-tune of Sarvam 105B that scores supplied options in one forward pass rather than generating a response token by token.
Sarvam reports an average score of 71.8 across three benchmarks and a 44-millisecond response time; it says the model ranked highest among 25 tested models with response times below 100 milliseconds.
The weights are available on Hugging Face under Apache 2.0, and Sarvam says local use requires a single GPU with at least 120 GB of memory.

Summary written by Lux AI from 2 articles

Key facts

Model: Prism is a LoRA fine-tune built on Sarvam 105B.
Reported response time: 44 milliseconds.
Reported average benchmark score: 71.8 across three benchmarks.
Licence: Apache 2.0.
Local hardware requirement: A single GPU with at least 120 GB of memory, according to Sarvam.