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

Miranda, Lester James V.; Hu, Songbo; Reichart, Roi; Korhonen, Anna

Where and how language models are deployed determines who can benefit from them. Communities in the Global South face what we call the last mile: the intersection of multilinguality and edge deployment, where the goals are aligned but the technical requirements often compete. We survey 306 papers that tackle this problem across the language modelling pipeline, from data collection to development and deployment, and provide actionable recommendations for different stakeholders in the NLP ecosystem.

Read the full abstract

Where and how language models (LMs) are deployed determines who can benefit from them. However, there are several challenges that prevent effective deployment of LMs in non-English-speaking and hardware-constrained communities in the Global South. We call this challenge the last mile: the intersection of multilinguality and edge deployment, where the goals are aligned but with competing technical requirements. We define the last mile along the dimensions of digital infrastructure and language capability, and describe the requirements and constraints they impose for developing and deploying accessible language technologies. To understand the state of the art and the challenges at this intersection, we survey 306 papers from 2021 to 2026 that tackle this problem across the language modelling pipeline, from data collection to development and deployment. We also discuss open questions and provide actionable recommendations for different stakeholders in the NLP ecosystem. Finally, we hope that this work contributes to a renewed understanding of what inclusive and equitable language technologies require to deliver meaningful benefit to the communities they serve.

This website is better viewed on a desktop

What shapes multilingual edge LM development, and how can these models reach the communities that need them the most?

Communities with the highest linguistic diversity often face the most severe infrastructure constraints. Countries such as Papua New Guinea, Nigeria, and Chad are linguistically diverse yet among the least active on the Internet.

[ figure loads here ]

Figure 1. Countries with high linguistic diversity have the most limited network connectivity (upper left). Internet penetration is sourced from ITU (2025), number of living languages (log-scale) from the Ethnologue (SIL International, 2025), and income groups from the World Bank (2025). Hover a point for the country.

Answering this requires combining multilinguality and edge NLP. But combining the two areas is hard: while their goals are aligned, their technical requirements often compete, and the two research communities remain largely siloed. The field has several names for this tension: the low-resource double bind (Ahia et al., 2021), the square-one bias (Ruder et al., 2022), Zeno's paradox of language technology (Nigatu et al., 2024), among others.

We define this challenge of developing and deploying LMs in contexts where language capability is limited and digital infrastructure is severely constrained as the last mile.

Surveying the Last Mile

To understand the state of the art and the challenges at this intersection, we survey 306 papers1 that tackle this problem across the language modelling pipeline. Below, we show an overview of these papers based on the languages they support and the range of model sizes they release.

[ figure loads here ]

Figure 2a. Reported language coverage of edge LM papers. We show 102 papers (of 306) that report a concrete number of evaluated languages and bin them into four brackets: monolingual (1), few (2–10), many (11–50), and massive (50+), categorized by research focus.

[ figure loads here ]

Figure 2b. Model sizes (in billion parameters) of various language models. For each model family in our curated set of released models, we recorded publicly documented parameter counts and plotted the range of available sizes on a log scale.

We also describe four hardware classes across this landscape and distinguish deployment modes for each class, namely, on-device, where the model runs directly on the hardware; and client, where the device sends requests to a remote or local server.

Microcontrollers

ESP32, Arduino
512 KB–16 MB RAM
On-device (~30M)

Single-board computers

Raspberry Pi, NVIDIA Jetson
1–16 GB RAM
On-device (~3B)

Smartphones

Tecno Spark 20, Galaxy A16, iPhone 16 Pro
4–24 GB RAM
On-device (~3B, ~7B‑Q4); client

Consumer PCs

Chromebook, MacBook, RTX laptops
8–32 GB RAM
On-device (7B–32B); client

Figure 2c. Hardware classes for edge LM deployment. We organize edge devices into four classes according to their technical specifications. For Memory, we report volatile working memory which governs the size of the model that can be loaded. For Compute, we describe the processing throughput (in trillion operations per second (TOPS) or GHz) which dictates inference speed.

Methods for Building Edge Models

The requirements for deploying on the edge and supporting multilinguality often compete, imposing challenges across the language modelling pipeline. Click on each pipeline stage (or requirement) to read about the challenges and the state of the art.

[ pipeline diagram loads here ]

Deploying Edge Models in Practice

We also looked into edge LM systems, which we define as completed efforts that have been integrated into real-world applications. To identify them, we manually classified each of the 306 papers on whether an actual model deployment took place, obtaining 58 systems in the process.

How are edge LMs developed?

To examine how edge LM systems are made, we situate the 58 deployment papers within the broader 306 surveyed papers. We embed each abstract with MiniLM, reduce to 2D with UMAP, and cluster with HDBSCAN. Then, we use KeyBERT to extract the top keywords per cluster. Hover any cluster to see representative papers.

[ chart loads here ]

Figure 3. Clustering of the 306 surveyed papers by abstract similarity. Real-world deployments (★) tend to concentrate near a few clusters such as model compression and dialog datasets, while clusters like reasoning performance or prompt compression have little to no representation, suggesting that edge LM deployments favor a relatively narrow set of methods.

Who develops edge LMs?

We classified the affiliations of authors across the 58 deployment papers into four sectors: Academia (universities and affiliated research institutions), Industry (startups to enterprise), Research collective (non-profit research organizations), and Government (state-affiliated institutes, public sector). Authors with multiple affiliations are counted in each. Cross-sector collaborations are measured by how often each pair of sectors co-occurs within the same paper.

[ chart loads here ]

Figure 4. Affiliation type of authors from papers that deployed edge LM systems. Numbers on each arc show the total count of papers contributing from that sector. Academia has the largest proportion of collaborations, while government participation remains limited and is mostly driven by cross-sector ties with academia.

Which domains are edge LMs deployed to?

In order to map the domains in which an edge LM is deployed, we perform a round of classification by tagging each paper according to their domain: Agriculture, Climate, Finance, Healthcare, Legal, Social, and Speech. Then, we extract mentions of different methods by keyword matching via KeyBERT, and visualize the domain-method connections as a network graph. Click on any outer domain node to see representative papers for that domain.

[ chart loads here ]

Figure 5. Edge LM real-world deployment domains network. Central nodes represent methods used to develop and deploy real-world edge LMs. Edge color indicates connectivity, while darker nodes indicate high sharing among domains.

Recommendations

For NLP researchers and model developers: Evaluate edge models beyond memory (e.g., compute and energy), and explore underrepresented methods since current deployments cluster around a relatively narrow toolkit.
For deployment practitioners and communities at the edge: Build cross-sector collaborations (academia, industry, research collectives, government), and involve local communities as active collaborators in development and deployment.
For policymakers and funders: Invest not only in model development but also in infrastructure and devices that make deployment feasible in linguistically diverse, lower-resource settings; increase public-sector participation in edge LM efforts.

Citation

@misc{miranda2026multilingualityedgedevelopinglanguage,
  title={{M}ultilinguality at the {E}dge: {D}eveloping {L}anguage {M}odels for the {G}lobal {S}outh},
  author={Lester James Validad Miranda and Songbo Hu and Roi Reichart and Anna Korhonen},
  year={2026},
  eprint={2604.21637},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2604.21637},
}

Discussion

Have feedback, questions, or ideas? Join the conversation below.

Methodology

To understand the landscape of multilingual and edge LMs, we collect research from January 2021 to March 2026 on multilinguality, edge and efficient NLP, or both. In total, we obtain 306 papers.

306 papers
58 deployment papers

Identification of papers

Keyword search. We gather studies from *CL (e.g., ACL, EMNLP, TACL), machine learning venues (e.g., NeurIPS, ICLR, TMLR), related workshops, and preprints using the Semantic Scholar API. Our search query includes keywords pertaining to multilingual NLP, efficiency, and LM deployment. This step resulted in 2,473 papers.

Targeted search. To cover work outside mainstream *CL and ML venues, we perform a targeted search across three groups of venues: regional venues such as AfricaNLP and PACLIC, efficiency-focused workshops such as SustaiNLP and ENLSP, and human-computer interaction and education venues where deployments are commonly reported, such as CHI and AIED. We also include grey literature such as technical reports and model cards that describe deployed systems. Within these sources, we include papers that study a non-English language, a hardware constraint, or a real-world deployment. This step resulted in 55 papers.

Screening and annotation of papers

Citation filter. We screen the keyword search results by citation count in a staggered manner, with thresholds by publication year. We apply this filter to keep annotation tractable, as the keyword search returns thousands of papers. In addition, the targeted search results are exempted from this filter so as to accommodate relevant works that may be disadvantaged by this criterion. Combined, this step led to an initial set of 1,069 papers for annotation.

LM-assisted annotation. In order to characterize each work, we annotate each paper across several dimensions. We bootstrapped all annotations using GPT-4.1 Mini, prompting the model with each paper's title and abstract.

Pipeline stage (multi-label): which stage(s) of the LM pipeline the work contributes to: Data Collection, Pretraining, Post-training, Inference, Evaluation, or Full-Stack.
Focus (single-label): whether the work specifically targets multilinguality or edge efficiency, or both.
Deployment device and mode: the target hardware or device for deployment, if specified, and whether a work was deployed on-device or as a client.
Deployment (yes/no): whether an actual deployment has taken place, i.e., any mention of usage or community-based evaluation or feedback.
Artifacts: for papers that release a model or dataset, its name, parameter count (if model), and whether it was used for training or evaluation (if dataset).

Post-correction. We conduct a round of post-correction to verify the LM-bootstrapped annotations. We check whether the labels are correct given the title and abstract; if they appear incorrect or ambiguous, then we consult the full paper and make a correction. During this step, we also exclude papers that (1) address neither multilinguality nor efficiency, (2) do not involve language as a modality (e.g., vision-only work), or (3) have no technical contribution such as surveys and position papers. This reduces the set to 306 papers.

Included papers in the survey

These 306 works range from peer-reviewed publications to workshops, regional venues, and grey literature. Out of the 306 papers, 6.9% are model releases, 74.2% describe methodologies for developing these edge LMs, while the rest (58 papers, 19.0%) are real-world system deployments.