💬 NLP

Multilinguality at the Edge: Developing Language Models for the Global South

This paper addresses the "last mile" challenge of deploying language models in the Global South by surveying 232 studies on the intersection of multilingual NLP and edge computing, offering actionable recommendations to overcome technical constraints and promote equitable language technologies.

Original authors: Lester James V. Miranda, Songbo Hu, Roi Reichart, Anna Korhonen

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lester James V. Miranda, Songbo Hu, Roi Reichart, Anna Korhonen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, all-knowing librarian who speaks every language in the world. This librarian is a Large Language Model (LM). They can write stories, solve math problems, and give advice in hundreds of tongues.

However, there's a catch. This librarian lives in a massive, high-tech palace (the Cloud) with unlimited electricity and super-fast elevators. To talk to them, you need a golden ticket: a fast internet connection and a powerful smartphone.

Now, imagine a village in the "Global South" (parts of Africa, Asia, and Latin America). Here, people speak hundreds of unique, local languages. But their internet is slow or non-existent, and their phones are old, small, and have tiny batteries. They can't afford the golden ticket, and the librarian in the palace can't hear them.

This paper is about building a traveling librarian who can live inside these small, old phones, speak the local languages fluently, and work without needing a super-fast internet connection. The authors call this journey the "Last Mile."

Here is the story of their journey, broken down simply:

1. The Two Opposing Forces

The authors realized that trying to build this traveling librarian is like trying to fit a whale into a teacup.

The Whale (Multilingual Capability): To speak 100+ languages well, the librarian needs a huge brain (lots of data and memory).
The Teacup (Edge Constraints): The phones in these villages are small. They have limited memory, weak processors, and tiny batteries. If the librarian is too big, the phone crashes. If the librarian is too small, they forget how to speak the local languages.

Usually, scientists study these two problems separately. Some try to make the whale smaller; others try to teach the whale more languages. But this paper says: "We need to study them together!" because the people who need the librarian the most face both problems at once.

2. The Journey: Five Stops on the Pipeline

The authors looked at 232 research papers to see how people are trying to solve this puzzle. They broke the process down into five stops:

Stop 1: Gathering the Books (Data Collection)
- The Problem: The librarian needs books in local languages to learn. But there are very few books written in these languages online.
- The Fix: Instead of just waiting for books, researchers are using AI to write new books (synthetic data) or carefully cleaning up messy internet text to make sure the librarian learns the right words.
Stop 2: The School (Pretraining)
- The Problem: Teaching the librarian takes a massive amount of energy and time.
- The Fix: Researchers are designing smarter "classrooms." They use special techniques to teach the librarian efficiently, so they don't need a super-computer to learn. They also use "shared vocabulary," where learning one language helps the librarian learn another, saving space.
Stop 3: The Specialization (Post-training)
- The Problem: The librarian is smart but maybe too big to fit in the phone.
- The Fix: This is like compression. Imagine taking a giant encyclopedia and shrinking it down to a pocket-sized guide without losing the important facts. Techniques like "quantization" (making the numbers smaller) and "distillation" (having a big teacher teach a small student) help shrink the librarian's brain.
Stop 4: The Conversation (Inference)
- The Problem: When the librarian talks, it uses up the phone's battery. Some languages are "expensive" to speak because the phone has to break them into tiny pieces (tokens) inefficiently.
- The Fix: Researchers are teaching the librarian to speak faster and use less energy. They use tricks like "speculative decoding" (guessing the next word before saying it) to speed things up, and they optimize how the phone reads the text.
Stop 5: The Report Card (Evaluation)
- The Problem: How do you know the librarian is actually good? Testing them on every language takes forever and costs too much money.
- The Fix: Instead of testing everything, they are creating "Lite Benchmarks"—short, smart quizzes that tell you if the librarian is ready for the village without needing a full marathon test.

3. Who is Building This?

The paper found that the people building these traveling librarians are mostly universities and research groups. They are often working together with NGOs (non-profits) and tech companies.

Who is missing? Governments are rarely involved. The authors say we need more government support to build the roads (infrastructure) and buy the phones so these tools can actually reach the people.

4. The Big Picture: Why It Matters

The authors argue that if we don't solve this "Last Mile" problem, the gap between the rich world (Global North) and the rest of the world (Global South) will get wider.

Without this: Only people with fast internet and expensive phones get to use AI.
With this: A farmer in Kenya, a doctor in Rwanda, or a teacher in India can use AI on their basic phone to get medical advice, learn new skills, or manage their crops, all in their own language.

The Takeaway

This paper is a roadmap. It tells us that we can't just make AI smaller, and we can't just make AI speak more languages. We have to do both at the same time, with a deep respect for the people who will use it.

It's about taking the "Whale" and teaching it to swim in the "Teacup," so that everyone, no matter where they live or how much money they have, can benefit from the magic of language technology.

1. Problem Statement: The "Last Mile" Challenge

The paper identifies a critical gap in the deployment of Large Language Models (LLMs): while models are becoming more multilingual and efficient, they remain inaccessible to communities in the Global South. These communities face a "double bind":

Linguistic Diversity: They speak thousands of low-resource languages often ignored by standard models.
Infrastructure Constraints: They suffer from limited internet connectivity, low smartphone penetration, and hardware constraints (memory, compute, energy).

The authors define this intersection as the "Last Mile": the challenge of deploying models that are both multilingual (understanding diverse, low-resource languages) and edge-ready (running on constrained hardware without cloud dependency). The core technical conflict is that the requirements for multilingual capability (large parameter counts, diverse data, complex tokenization) often compete with the requirements for edge deployment (small model size, low latency, low energy consumption).

2. Methodology

To understand the state of the art and the tensions between these competing requirements, the authors conducted a comprehensive survey of 232 research papers published from 2021 onwards.

Data Collection: Papers were sourced from top-tier NLP (ACL, EMNLP, TACL) and ML (NeurIPS, ICLR, TMLR) venues using Semantic Scholar.
Screening Criteria: Papers were filtered based on relevance to multilinguality, efficiency, and edge deployment. A staggered citation filter was applied (e.g., 2021 papers required $\ge$ 100 citations; 2025 papers required $\ge$ 3).
Annotation Process: A hybrid human-AI approach was used. GPT-4.1 Mini initially annotated papers based on title/abstract against specific dimensions (pipeline stage, focus on edge/multilinguality, model size, languages supported), followed by rigorous human validation.
Framework: The survey organizes the literature along the Language Modeling (LM) Pipeline:
1. Data Collection
2. Pretraining
3. Post-training (Fine-tuning, Compression, Merging)
4. Inference
5. Evaluation

3. Key Contributions & Findings

A. The Pipeline Tensions (Section 4.2)

The paper details how edge constraints and multilingual requirements clash at every stage of the LM pipeline:

Data Collection:
- Challenge: Limited model capacity leads to the "curse of multilinguality" (adding languages dilutes performance) and susceptibility to noise in low-resource data.
- Methods: Vocabulary sharing, synthetic data generation (using strong teacher models), and rigorous data curation/filtering.
Pretraining:
- Challenge: Multilingual performance scales with model size, but edge devices cannot host large models. Additionally, multilingual complexity strains edge compute.
- Methods: Optimized tokenizer design (subword allocation, byte-level encoding), continual pretraining (vocabulary expansion without full retraining), and parameter-efficient architectures (LoRA, Adapters, Mixture-of-Experts).
Post-training:
- Challenge: Base models are often too large for device memory; adding/removing capabilities requires expensive retraining.
- Methods: Model compression (quantization, pruning), Knowledge Distillation (off-policy and on-policy), Model Merging (combining weights of different models), and Machine Unlearning (removing sensitive data without retraining).
Inference:
- Challenge: Inference costs are non-uniform across languages due to tokenization inefficiencies ("token tax"), draining battery and increasing latency.
- Methods: Prompt compression, Speculative Decoding (using small draft models), and inference-time adaptation (vocabulary expansion without retraining).
Evaluation:
- Challenge: Evaluating multilingual models is computationally expensive and benchmarks are often outdated.
- Methods: "Lite" benchmarks, instance selection (using IRT/RL), and dynamic evaluation generation.

B. Analysis of Real-World Systems (Section 5)

The authors analyzed 36 papers describing actual edge LM system deployments (e.g., healthcare chatbots in India/Rwanda, agricultural tools in Kenya).

Development Patterns: Deployments heavily rely on a narrow set of methods: Model Compression (quantization) and Synthetic Data. There is a lack of experimentation with other advanced techniques like speculative decoding or complex orchestration.
Stakeholders: Successful deployments are rarely single-entity efforts. They predominantly involve cross-sector collaborations (Academia + Research Collectives + Industry). Government participation is currently limited.
Domains: Deployments are most common in Healthcare and Social sectors, followed by Agriculture and Speech.

C. Emerging Trends

Model Sizes: There is a shift toward "Small" models ( $\le$ 8B parameters) suitable for mobile devices, with families like Qwen2.5 and Gemma 3 offering variants from <1B to >30B.
Language Coverage: While multilingual papers cover 50+ languages, "Edge-only" papers are overwhelmingly monolingual. The intersection of both edge and multilinguality remains underexplored.

4. Significance and Recommendations

The paper argues that solving the "Last Mile" requires moving beyond siloed research. It provides actionable recommendations for different stakeholders:

For Researchers: Evaluation of edge models must consider energy and compute constraints, not just parameter size. Researchers should explore underrepresented methods (e.g., beyond quantization) to find better trade-offs.
For Practitioners & Communities: Successful deployment requires partnerships between research collectives, NGOs, and industry. Communities should move from being "consultants" to active collaborators in the development process to ensure sovereignty and relevance.
For Policymakers: Funding should not only target model development but also the infrastructure and devices required to run them. Increased government involvement is necessary to bridge the gap.

5. Conclusion

The paper concludes that reaching the last mile is a continuous process, not a one-time fix. It calls for a holistic approach that integrates multilingual capability with edge efficiency at every stage of the LM pipeline. By addressing the "low-resource double bind" (data scarcity + compute scarcity) through collaborative, context-aware development, the field can move toward more equitable and inclusive language technologies for the Global South.

Key Takeaway: The technical requirements for multilingualism and edge deployment often compete, but through specific pipeline optimizations (tokenization, architecture, compression) and cross-sector collaboration, it is possible to build inclusive, on-device language models for the world's most linguistically diverse and resource-constrained communities.