SquadStack Maya is a new large language model built specifically for sales conversations, launched by SquadStack.ai in November 2026. The model was post-trained on more than 40 crore (400 million) real sales interactions collected over six years, including how each conversation ended.
Maya is built on an open-weight foundation model, though SquadStack has not disclosed which model, its size or its full training method, CIOL reported on 16 November 2026. Its training data records whether each conversation ended in a sale, a callback or a drop-off. Maya already runs inside SquadStack’s enterprise Voice AI Agents platform, and new enterprise customers can request beta access. It supports Hinglish and Indian languages, and SquadStack says the model is developed and hosted in India, with conversations processed domestically.
How Does SquadStack Maya Perform Against Other AI Models?
SquadStack tested Maya on its own Telesales LLM Benchmark, built from 9,962 conversation moments taken from 471 real Indian sales calls. Eleven models generated the next agent reply in each situation, and the replies were scored on seven checks, including remembering context, following campaign instructions and using tools correctly. Anthropic’s Claude Haiku 5.5 acted as the automated judge.
- Maya: 96.7% pass rate
- Google Gemma 4 31B: 92.5%
- DeepSeek V4.1 Flash: 92.2%
- OpenAI GPT-6 Luna: 89%
Maya kept context on 96% of turns, against 90% for the next best model, and chose the appropriate reply 95% of the time, against 87% for its closest competitor.
On speed, SquadStack says Maya’s median time to start a reply is 278 milliseconds on its production systems in India, and 95% of replies begin within 774 milliseconds. Gemma 4 31B took about 2.3 seconds at the same 95th-percentile mark. The models were not tested under identical hosting conditions, and the figures exclude speech recognition, audio generation and network delays.
These results are self-reported. The benchmark was built by SquadStack, scored by another AI model, and measures reply quality rather than actual sales conversions.
What It Means for Indian Businesses
SquadStack handles more than 50 lakh (5 million) conversations a day on its voice AI platform for customers in lending, insurance, education and commerce. According to co-founder and CTO Vikas Gulati, Maya has processed more than five trillion tokens and handles over one trillion tokens a month.
For Indian sales teams, the launch reflects a shift from general-purpose AI models to models trained on one specific job and local language patterns. SquadStack has not yet shown that the benchmark gains lead to more completed sales or higher revenue.
Industry Reaction and Expert Commentary
Gulati said: “A live sales call is the hardest place to run an LLM,” adding that “Thousands of calls can start in the same second, and the customer hears every slow turn.”
Co-founder and CEO Apurv Agrawal said: “Anyone can download a model. Nobody can download six years of real sales conversations and what happened after each one.”
No independent expert reaction had been published at the time of writing.
What Happens Next?
SquadStack plans to add the ability to read pauses, tone changes and information from earlier customer interactions, though these features are not part of the launch. The company also develops Arth V2, a speech-recognition model for Indian sales conversations.
Frequently Asked Questions
What is SquadStack Maya?
It is a large language model built for sales conversations, post-trained on more than 40 crore real sales interactions.
Does Maya support Indian languages?
Yes. SquadStack says Maya supports Hinglish and Indian languages and is hosted in India.
Can businesses use Maya now?
Maya runs inside SquadStack’s Voice AI Agents platform, and new enterprise customers can request beta access.
Source: CIOL, 16 November 2026.
Leave a comment