Overview
The survey compiles many practical, open datasets and compact models, making real deployments feasible, but coverage and standardization remain uneven across Indic languages.
Citations2
Evidence Strength0.78
Confidence0.80
Risk Signals8
Trust Signals
Findings with numeric evidence: 6/6
Findings with evidence refs: 6/6
Results with explicit delta: 1/4
Reproducibility
Status: Partial assets available
Open source: Partial
At A Glance
Cost impact: 65%
Production readiness: 60%
Novelty: 45%
Why It Matters For Business
Indic languages cover ~1.5–2 billion speakers; focused datasets and compact models enable local-language products with much lower compute and cost than retraining huge universal models.
Who Should Care
Summary TLDR
This is a focused literature review of 84 recent papers on large language model (LLM) work for Indic languages. It catalogs models, corpora, benchmarks, tools, and methods; highlights uptake of small efficient models, large parallel corpora, and many new benchmarks; and lists recurring gaps: data scarcity, script and dialect diversity, code-mixing, and inconsistent evaluation. The paper is a practical map for building or evaluating Indic NLP systems and for prioritizing dataset and benchmark work.
Problem Statement
Indic languages cover huge linguistic diversity but are under-served by LLM research. Progress is patchy due to limited high-quality data, many scripts and dialects, frequent code-mixing, and few standardized benchmarks. This review aims to summarize what exists, surface gaps, and offer a taxonomy to guide researchers and practitioners.
Main Contribution
Systematic review and curated bibliography of 84 recent papers on Indic LLM research.
A taxonomy that groups work into five areas: LLMs, Corpora, Benchmarks & Evaluation, Techniques, and Tools & Applications.
Key Findings
Number of papers reviewed
Small efficient Indic models can match or beat larger general models on local-language generation
Results
| Metric | Value | Baseline | Delta | Split / Dataset | Evidence | Evidence Ref |
|---|---|---|---|---|---|---|
| papers reviewed | 84 | — | — | survey corpus | Authors screened >1000 titles/abstracts and included 84 studies | Section 2.3 |
| Paramanu model sizes | 13.29M–367.5M params; context 1024 | — | — | Paramanu paper summary | Paramanu family reported model sizes and context length | Section 3.1.1 |
What To Try In 7 Days
Inventory which Indic languages and scripts your product needs and map to Samanantar, Vistaar, or language-specific corpora.
Test a small, language-specific model (e.g., Paramanu or instruction-fine-tuned Airavata/Tamil-LLaMA) before defaulting to large general LLMs.
Prototype PEFT adapters (LoRA/FLix) on an open model to measure gains with minimal compute.
Optimization Features
Token Efficiency
Model Optimization
Training Optimization
Reproducibility
Data URLs
Risks & Boundaries
Limitations
Survey depends on included 84 papers; rapidly evolving field may have new work beyond their cutoff.
Heterogeneous reporting in source papers prevents uniform quantitative comparison.
When Not To Use
When you need certified high-stakes accuracy in under-represented dialects without further validation.
When a specific Indic script or dialect was not covered by the referenced datasets or models.
Failure Modes
Tokenization inequality: some scripts use many more tokens, raising cost and shrinking context.
Cross-lingual contamination or benchmark leakage may inflate reported scores.