1 comment

[ 4.8 ms ] story [ 16.8 ms ] thread
A new AI inference startup, Avartha, delivers the fastest, most accurate and affordable frontier-grade voice AI using open weights models.

It beats OpenAI, Google, xAI, and Phonic on τ³-Voice task completion.

Avartha’s optimized inference stack supports up to 10x more concurrent conversations per GPU than SOTA open-source inference engines vLLM and SGLang.

Spun out of Georgia Tech, Avartha’s first release, Ava Zero, achieves 86% task completion at just $2.50 per conversation hour.