High-quality data pipelines
Curation, filtering, and synthesis for Indic and Global South text, code, math, and speech.
// PROJECT EKA
Backed by the IndiaAI Mission
Sovereign models for reasoning & multilingual AI
Project EKA is Soket's flagship research program — advancing frontier math, code, and reasoning models in parallel with sovereign multilingual AI for Indian and Global South languages.
1,536
H100 GPUs
Sovereign compute via IndiaAI Mission
25T
Tokens curated
High-quality pre-training corpus
120B+
Parameters
Frontier-scale architecture in training
60+
Languages
Indian, Global South, and programming
// WHAT WE'RE BUILDING
Foundation models built for Bharat & the Global South
Project EKA is Soket's boldest vision — building AI for a billion, from the heart of India. Our mission is to create world-class models that master math, code, and reasoning, while speaking the languages of Bharat and the Global South.
We work at the edge of research in architecture, large-scale training, and language resources — reimagining what's possible for low-resource and diverse languages.
We open-source where we can, train on sovereign compute, and publish research that advances Indic NLP, systems for ML, and efficient inference.
// CORE CAPABILITIES
Two tracks: technical reasoning & multilingual AI
01
Math
Advanced mathematical reasoning, proofs, and symbolic computation — models tuned for rigorous step-by-step logic.
02
Code
Multi-language code generation, debugging, and optimization across 20+ programming languages.
03
Reasoning
Logical deduction, complex analysis, and long-horizon problem solving for high-stakes workflows.
04
Multilingual
22 Indian languages, 20+ Global South languages, and English — sovereign text and speech with curated data and tokenization.
// LANGUAGE COVERAGE
60+ languages — not an afterthought
Most frontier labs optimize for English. EKA runs two parallel verticals: frontier math, code, and reasoning models for rigorous technical work — alongside sovereign multilingual modeling with efficient tokenization and curated corpora.
22 Indian languages
Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, Assamese, Urdu, Sanskrit
20+ Global South languages
Arabic, Indonesian, Thai, Vietnamese, Burmese, Kazakh, Portuguese, Spanish
20+ programming languages
Python, Rust, Go, TypeScript, C++, Java, SQL, Julia
// RESEARCH DIRECTIONS
Problems we're actively working on
If you care about data systems, training at scale, tokenizers, post-training, or ethical AI — these are the threads where your work ships into a national-scale model.
01 02 Efficient model architecture
Routing strategies for token-efficiency and morphological diversity across scripts.
03 Efficient tokenization
Token-efficient vocabularies for Indian languages — minimizing bytes-per-token for Indic scripts.
04 Post-training methods
SFT and preference optimization for math, code, and reasoning; alignment for regulated sectors.
05 Efficient inference
Kernel fusion, speculative decoding, quantization, and systems co-design.
06 Sustainable AI research
Optimal power and water usage — minimizing environmental cost of frontier runs.
07 AI for critical sectors
Defence, cybersecurity, finance, banking — with auditable, on-premise deployment.
08 Ethical AI
Safety, alignment, bias mitigation — embedding constraints from data to deployment.
// RELATED RELEASES
EKA Tokenizer
Token-efficient vocabularies for Indian and Global South languages.
Help us train India's frontier models
We're hiring researchers and engineers across data, training, inference, and applied ML. If you want hard systems problems at sovereign scale — we'd like to hear from you.