Skip to content

Sarvam open-sources Sarvam Prism, a System 1 model built on Sarvam 105B that picks answers from given options with a 44 ms median response

Tue, Nov 3 22:16 UTC+8

Pratyush Kumar@pratykumar

Open-sourcing our System 1 model Sarvam Prism, trained on Sarvam 105B. It pushes the pareto of accuracy v speed v calibration. Uses block-bidirectional attention, isolates options with a mask and shared positions, and a split attention kernel for fast inferencing. x.com/SarvamAI/status/211762141…

XTue, Nov 3 22:16 UTC+8

Sources1

@pratykumarXTue, Nov 3 22:16 UTC+8Open-sourcing our System 1 model Sarvam Prism, trained on Sarvam 105B. It pushes the pareto of accuracy v speed v calibration. Uses block-bidirectional attention, isolates options with a mask and shared positions, and a split attention kernel for fast inferencing. https://x.com/SarvamAI/status/2117621415520448838View original · Tue, Nov 3 22:16 UTC+8

Headline by dalao.ai. The quoted post is the record; its author keeps authorship.