2 Sources
[1]
SquadStack.ai launches Maya, claims its Sales LLM outperforms models from OpenAI, Google
SquadStack.ai has introduced Maya, a Sales Large Language Model designed to excel in sales conversations. Maya achieved a remarkable score of 96.7% in internal evaluations, surpassing other AI models. The model is trained on over 40 crore real sales interactions, enhancing its understanding of
[2]
SquadStack.ai launches Maya, the world's first Sales LLM, beating OpenAI, Google and leading open-source models on real sales calls
Maya passed 96.7% of quality checks on 471 real Indian sales calls and began its reply in 278 milliseconds, less than half the time of the closest model on quality. Post-trained on more than 40 crore real sales interactions and their outcomes, it already runs over a trillion tokens a
Share
Copy Link
SquadStack.ai launched its sales AI Maya, saying it beat OpenAI and Google models on a benchmark the company built itself.
SquadStack.ai, an Indian AI startup that has spent six years running sales call operations, launched Maya on November 16, a model it claims is the world's first Sales Large Language Model
2
1
. It says Maya outscored models from OpenAI, Google and DeepSeek on a sales benchmark the company itself designed and ran2
1
.
Source: ET
The pitch starts with a problem the company says nobody has solved. The company says live sales calls are where generic models fail. An agent cannot tell a hesitant "haan theek" from a real yes, it argues, and nearly half the time replies to speech it only half heard. No prior model was trained on these moments, it says
2
.On its own benchmark, Maya scored 96.7%
2
1
. Google's Gemma 4 31B followed at 92.5%2
1
. DeepSeek V4.1 Flash managed 92.2%2
1
. OpenAI's GPT-6 Luna, the weakest of the named rivals, trailed at 89%2
1
. The test put 11 models through 9,962 conversational moments pulled from 471 real Indian sales calls, with Anthropic's Claude Haiku 5.5 scoring each reply pass or fail on seven checks2
1
.The sub-scores give the flavour. Maya kept track of the conversation on 96% of turns, per the company's benchmark, with the next best model at 90%
2
. It broke campaign rules on 4% of replies, which the same benchmark puts at less than half Gemma 4 31B's rate2
. GPT-6 Luna failed 30% of its tool-use checks and sometimes hung up on customers mid-sentence2
.Speed is the second pillar of the launch claim. The company says replies start in a median 278 milliseconds on its own production endpoint in India
2
. Gemma 4 31B's comparable figure, at 2.3 seconds, was measured through a shared hosting provider, per the company2
. The models that matched Maya's speed on similar dedicated setups passed only 66 to 72% of the benchmark's checks2
.The benchmark behind the launch is the company's own, built from its own data and judged on criteria it chose
1
. It is published on Hugging Face with open submission, so outside models can be run against it2
.The company has run sales call centres for six years, and those calls, tagged with what happened after each one, are what Maya learned from. It was post-trained, starting from an open-weight foundation model, on over 400 million outcome-tagged interactions, each marked as a sale, a callback or a drop-off
2
1
.This is not a demo. Maya already handles over five million conversations daily in Hinglish and other Indian languages, the company says, and has processed over five trillion tokens in total
2
1
. It runs inside SquadStack.ai's Voice AI Agents for enterprise customers; new brands can request beta access2
. Because it is built and hosted in India, conversations stay in-country, which the company says simplifies compliance with India's Digital Personal Data Protection Act for banks, lenders and insurers2
1
.Maya is the second model from the company's in-house operation, after the speech recognition model Arth V2
2
1
. A system named Lift reviews live conversations and fixes errors each week, the company says, with reading of pauses and tone changes planned next2
1
.Co-founder and CEO Apurv Agrawal argues the next frontier of models will come from companies that own a single domain deeply
2
. SquadStack.ai, in his telling, holds that position for sales in India. As he put it, "Anyone can download a model. Nobody can download six years of real sales conversations and what happened after each one." Whether the scores survive testing from outside, the open benchmark may eventually show2
.Related Stories
Summarized by
Navi
16 Jun 2026•Startups

18 Feb 2026•Technology

09 Feb 2026•Technology
