FuturixAIAccuracy (%)
Higher is better · hover any bar
Introducing Shivaay
Intelligence without boundaries.
Meet Shivaay, FuturixAI’s 4-billion-parameter language model, built by FuturixAI as its own foundation model to turn complex questions into clear answers. Designed for developers, reasoning and Indian languages.
From a single prompt to a worked answer. Shivaay reasons through problems step by step and works in your teams’ languages, from Hindi to Tamil and Telugu.
- 4B
- Parameters
- 8
- NVIDIA A100 GPUs used to train it
- 91.0%
- ARC-Challenge, graded by FuturixAI
- 90.6%
- GSM8K, graded by FuturixAI
Benchmarks
Ahead of larger open models.
On two public benchmarks.
Shivaay at 4B scores above Gemma 2 (9B), Mistral NeMo (12B), Phi-3.5-mini, Llama 3.1 (8B) and Mistral (7B) on ARC-Challenge and GSM8K. Shivaay’s scores are not independent; see how we measured.
01ARC-Challenge
91.0%
Science multiple choice, 1,172 questions
02GSM8K
90.6%
Maths word problems, 1,319 questions
03Parameters
4B
Small enough for a single GPU
04Training GPUs
8
NVIDIA A100s, per FuturixAI’s CEO
Methodology
How we measured
Accuracy is reported on two public benchmarks. ARC-Challenge tests grade-school science reasoning with multiple-choice questions written to defeat simple retrieval. GSM8K tests multi-step arithmetic word problems, scored on the final answer.
Shivaay scores 91.04% on ARC-Challenge (1,067 of 1,172) and 90.6% on GSM8K (1,195 of 1,319), graded by FuturixAI with an answer parser that reads every response, using numeric match on GSM8K.
Scores for the other models come from the benchmark table in Microsoft’s Phi-3.5-mini-instruct model card (ARC-Challenge 10-shot, GSM8K 8-shot with chain of thought). Microsoft ran those with its own setup, which differs from Shivaay’s, so the comparison is indicative.
Read these with care
Shivaay was evaluated through the hosted API, in a single run, at a sampling temperature of 0.7 to 0.8, with prompts that differ from standard leaderboard harnesses. Each score comes from one run, with no repeats, so it should not be ranked against published leaderboards.
The intelligence layer
behind real work.
01
Reasoning
Plans and works through multi-step problems before it answers.
02
Science questions
Answers grade-school science questions written to defeat simple retrieval.
03
Built in India
Built by FuturixAI in Noida and trained on eight NVIDIA A100 GPUs.
04
Multilingual
Understands and responds in Hindi, Tamil, Telugu and other Indic languages.
05
Compact
At 4 billion parameters, small enough to serve on a single GPU.
Developers
Check the numbers
yourself.
Shivaay’s evaluation outputs are public on GitHub. Clone them and re-run the grading.
git clone https://github.com/FuturixAI-and-Quantum-Works/Shivaay_ARC-C
cd Shivaay_ARC-C
# 1,172 ARC-Challenge questions with Shivaay’s answers
# score: 91.04%git clone https://github.com/FuturixAI-and-Quantum-Works/Shivaay_GSM8K
cd Shivaay_GSM8K
# 1,319 GSM8K problems with Shivaay’s answers
# score: 90.6%Model card
What Shivaay is
Origin. FuturixAI’s own foundation model, built by FuturixAI (FuturixAI, November 2026).
Training data. Digit reports techniques inspired by Meta’s LIMA paper, which emphasise stylised prompts over massive datasets. Digit also quotes the developers saying stylised prompts and synthetic datasets annotated by GPT-4 helped optimise training. AIM adds GATE and IIT exam questions. (Digit, Analytics India Magazine)
Languages. Several Indic languages, including Hindi, Tamil and Telugu. (Digit)
In the press
Digit · Analytics India Magazine · Press release (syndicated via ANI)
Evaluation outputs: Shivaay_ARC-C · Shivaay_GSM8K
Build what comes next
with Shivaay.
FAQ
Questions, answered.
What is Shivaay?
Shivaay is a 4-billion-parameter foundation language model built by FuturixAI.
Is Shivaay a foundation model?
Yes. Shivaay is FuturixAI’s own foundation model, built by FuturixAI.
What hardware was Shivaay trained on?
FuturixAI’s CEO told Digit that the team trained Shivaay on eight NVIDIA A100 GPUs.
How does Shivaay score on benchmarks?
91.04% on ARC-Challenge and 90.6% on GSM8K, graded by FuturixAI in November 2026 from Shivaay’s published evaluation outputs.