FTA Global has introduced FTA Mark-1, its first specialist artificial intelligence model, at Bharat AI Innovation 2026 in Mumbai.
Developed using India’s open Sarvam-30B model and the company’s proprietary Micro-Judgement Tuning™ (MJT) methodology, FTA Mark-1 recorded winning results in separate chess challenges against OpenAI’s GPT-6 Astra, Google Gemini 4 Argon and Gemini 3.8 Flash.
The company selected chess as the model’s first application because the game offers a structured way to evaluate decision-making. Each move involves assessing possible options, anticipating outcomes and selecting an appropriate course of action. FTA Global used this environment to demonstrate how its MJT methodology can develop specialised judgement capabilities in an existing AI model.
In three independently organised 12-game challenges against models listed among the top four on Kaggle’s chess leaderboard as of 12 November 2026, FTA Mark-1 scored 10–2 against OpenAI’s GPT-6 Astra, with nine wins, two draws and one loss. Against Google’s Gemini 4 Argon, it recorded a 9–3 result, securing nine wins and three losses. Its challenge against Gemini 3.8 Flash also ended 9–3, comprising eight wins, two draws and two losses.
Across the 36 games, FTA Mark-1 recorded 26 wins, four draws and six losses, earning 28 points. FTA Global clarified that all challenges were organised and conducted independently by the company, with OpenAI, Google and Kaggle neither participating in nor endorsing the matches.
At the centre of the model is Micro-Judgement Tuning™, a training methodology developed by FTA Global. Building on published research, including approaches that score individual possible moves, MJT focuses on teaching AI models to evaluate different options, compare their relative merits and identify stronger decisions rather than concentrating primarily on reproducing the correct answer.
Instead of developing a new model from the ground up, the methodology works on top of an existing open model to build specialised capabilities for a particular task. FTA Global aims to make this approach more efficient in terms of training data, computing resources and development time.
FTA Mark-1 uses Sarvam-30B as its base model, supported by a lightweight Low-Rank Adaptation (LoRA) layer trained on positions from publicly available Lichess games. These positions were scored by the Stockfish chess engine during training. The model operates on a single 80 GB GPU, such as an NVIDIA H200 or A100.
During gameplay, FTA Mark-1 does not rely on an external chess engine. Its judgement model works alongside FTA’s think-ahead code, released openly with the model, to evaluate hundreds of smaller questions around possible moves before selecting a decision. The model also selects only from legal moves in each position, ensuring that its moves comply with chess rules. In match mode, it takes approximately one minute per move on average.
Speaking on the launch, Senthil Kumar Hariram, Founder & MD, FTA Global, said, “The next phase of AI will not be defined only by how large a model is, but by how effectively it can reason, evaluate choices and make decisions. With Micro-Judgement Tuning™, we wanted to explore a different approach to AI development, one where an existing open model can be trained to develop deep expertise in a specific task without requiring enormous amounts of data and computing power. Mark-1 is our opening step in demonstrating this possibility. More importantly, it shows that India can contribute to the global AI ecosystem not just by building large models, but by developing smarter and more efficient ways to train them.”
Senthil Kumar Hariram added, “We believe India has an opportunity to shape the next chapter of AI through efficiency and specialization. The objective is not necessarily to build the largest models, but to develop intelligent systems that can solve specific problems exceptionally well. With the Mark series, we want to explore how specialist AI can augment human expertise across industries, keeping people at the center of decision-making.”
FTA Global has also outlined the testing process to enable independent examination of the results. All participating models received identical instructions and operated under the same rule-checking framework. Each 12-game challenge used six fixed and balanced opening positions, with every opening played once with white and once with black. No chess clock was used, and while FTA Mark-1 operated with its own think-ahead mechanism, competing models ran at their makers’ default settings.
The company said the match records were timestamped on the Bitcoin blockchain before the results were announced. All 36 games, including individual moves and thinking times, have been published on Hugging Face and YouTube for researchers, developers and chess enthusiasts to review. A technical report on FTA Global’s website also details how frequently each model selected the best available move.
FTA Mark-1 is the first model in the company’s Mark series, which focuses on developing specialist AI capabilities. While chess serves as the initial application of MJT, FTA Global plans to extend the methodology to other areas where evaluating choices and making sound decisions are central to performance.
The company’s next area of focus is marketing intelligence. From January 2027, FTA Global plans to apply learnings from Mark-1 to client projects, developing systems that can assess multiple possibilities and help marketing professionals identify more effective strategies. Its broader Specialist Intelligence vision centres on AI models that enhance human expertise while keeping people involved in decision-making.
FTA Mark-1 is available on Hugging Face. Its adapter is released under the Apache 2.0 licence, the code under GPL-3.0 and the 36 match games under CC BY 4.0. The model is built on Sarvam-30B and is available for developers and researchers to download, test and play.
Send news announcements/press releases to:
jeevika@thefoundermedia.in
