TLDR
Many AI systems assume unlimited compute, cloud infrastructure, and connectivity. Those assumptions fail in much of the world.
Resource efficient AI requires alignment across the entire system: data governance, dataset design, model architecture, evaluation, and deployment.
Two Lelapa AI research contributions illustrate this approach:
- InkubaLM, a Small Language Model designed for efficient multilingual conversational AI
- The Esethu Framework, a governance model for building sustainable language datasets
Together they demonstrate how efficient AI emerges when data, models, and deployment environments are designed as one integrated system.
The Compute Constraint Shaping AI
The global AI race currently prioritises scale. Larger models, larger datasets, and increasingly powerful compute clusters dominate research and investment.
This trajectory assumes access to infrastructure that is far from universal.
Across many regions of the world, including large parts of the Global South, AI systems must operate under very different conditions:
- Limited GPU availability
- Unstable connectivity
- Constrained cloud access
- Devices with modest processing power
These realities change how AI must be designed. Systems that depend on massive compute resources struggle to operate reliably in these environments.
This challenge has accelerated interest in resource efficient AI, where systems are designed to perform well within real infrastructure constraints.
Efficiency is an Ecosystem
Efficient AI emerges from coordination across several layers of the AI stack.
These layers include:
- Data governance (defines how data is collected, owned, and managed over time)
- Dataset curation (ensures the data used for training is clean, relevant, and representative)
- Model architecture (determines how efficiently a model can learn and operate)
- Evaluation methods (measure how well models perform across real-world tasks)
- Deployment environments (shape how and where models are actually used in practice)
When these layers are misaligned, inefficiencies compound quickly, with large models trained on poorly curated datasets requiring more compute, weak governance making data difficult to maintain, and models designed for high-performance infrastructure struggling in constrained environments.
The result is technology that performs well in research settings but struggles in real-world deployment. At Lelapa AI, our approach focuses on designing these layers together, creating scalable language models and cost-efficient AI systems that perform reliably beyond research environments.
InkubaLM: Designing Language Models for Real Environments
InkubaLM is a Small Language Model (SLM) built for multilingual Natural Language Processing (NLP) in environments where compute resources are limited.
A language model is an artificial intelligence system trained to understand and generate human language. Many modern models contain billions of parameters, which are numerical values the system learns during training.
InkubaLM takes a more efficient architectural path. The model contains approximately 400 million parameters, compared to models that often exceed several billion. It was trained on 2.4 billion tokens of multilingual text, where tokens represent units of language such as words or fragments. By comparison, many large language models are trained on hundreds of billions to trillions of tokens, making InkubaLM far more data-efficient while still delivering strong performance.
This makes InkubaLM significantly lighter and more deployable across environments where traditional models are difficult to run. This positions InukbaLM among a new generation of lightweight AI models and efficient language models designed for practical use.
Despite its smaller size, InkubaLM performs strongly across several natural language tasks including sentiment analysis; natural language inference; question & answering; and translation.
This design allows the model to run across diverse environments including local servers, laptops, and smaller cloud infrastructure. The architectural philosophy is simple: AI models should match the environments where they will operate.
Compressing the Model Even Further
The Buzuzu Mavi Challenge, hosted by Lelapa AI in partnership with Zindi, explored how far InkubaLM could be optimised while maintaining performance.
More than 490 data scientists from 61 countries participated in compressing the model. Some submissions reduced the model size by up to 75 percent using various optimisation techniques.
These results show how AI model optimisation and compressed language models can be designed deliberately for efficiency and scale.
The Data Foundation: The Esethu Framework
Model architecture alone does not determine the efficiency or reliability of AI systems. The structure and governance of training data also play a critical role.
The Esethu Framework was developed to support sustainable language dataset development. Esethu introduces governance principles that guide how datasets are collected, curated, documented, and maintained. It also introduces mechanisms that ensure value created from language data contributes back to the ecosystems that produced it.
Key practices include:
- Transparent Dataset Documentation: Clear records describing dataset sources, composition, and intended use improve research transparency and reproducibility.
- Community Informed Data Creation: Language datasets benefit from participation by speakers, annotators, and linguistic experts who shape how data is collected and interpreted.
- Long Term Data Stewardship: Version control, defined ownership, and maintenance practices ensure datasets remain usable over time.
- Reinvestment in Language Data Ecosystems: Esethu also supports models where the value generated from language technologies helps fund the continued creation and maintenance of language datasets. This approach strengthens the long term sustainability of low resource language data and ensures that communities contributing linguistic knowledge remain part of the ecosystem that benefits from it.
Together, these practices produce stronger training foundations and more reliable and efficient AI systems.
A Blueprint for Resource Efficient AI
The AI industry is facing a clear constraint: models are growing faster than the infrastructure available to support them.
That constraint does not end at training. It extends into deployment, where the cost of inference, the ability to run models reliably, and the realities of infrastructure determine whether AI systems can operate at scale.
Models that are expensive to run or dependent on constant high-performance compute quickly become impractical outside these environments. This is where many systems fail, not at the model stage, but at the point of real-world use.
InkubaLM and the Esethu Framework show how to address this challenge through system design.
InkubaLM delivers efficient model architecture that reduces both training and inference costs. Esethu ensures sustainable, well-governed data that supports long-term model performance and maintainability. Together they form a blueprint for building resource efficient AI systems that scale across real-world environments.
As compute costs continue to rise and infrastructure gaps persist across many regions, this approach is becoming increasingly important across the global AI ecosystem.
At Lelapa AI, this is the foundation of how we build. Our work focuses on designing efficient AI models and scalable language models that perform in constrained environments. The core question shaping the industry is clear: how do we build AI systems that can scale beyond high-compute environments and into the real world?
At Lelapa AI, our answer is to design for efficiency from the start, across every layer of the system. That is the blueprint we are building.