🌍⚡ The Language–Energy Divide

Measuring Energy Costs of Multilingual LLM Inference

---

📢 News

• Aug 2026 — Our paper has been accepted to EMNLP 2026 (Main Conference)! 🎉

LLMs are increasingly deployed in multilingual settings, yet the energy cost of serving them across languages is poorly understood. We present the first systematic measurement of inference energy across 122 languages, and uncover a stark, persistent language–energy divide.

🔑 Key findings

	Finding
⚡	Per-token energy varies by up to 8.3× across languages (single model).
🔋	Total energy for a fixed request set varies by up to 179× — English (17.6 kJ) vs. Pashto (3,147 kJ).
📉	Double penalty: the most energy-expensive languages are also the least accurate.
🌐	The divide persists across 5 models, 2 GPUs, 6 batch sizes, and 3 tasks.

The disparity compounds two factors: higher per-token energy for complex/rare scripts, and more tokens generated for low-resource languages.

📊 What's in this repo

.
├── analysis/   # Scripts that build the paper's tables & figures, plus the
│               #   prompt-construction / NLLB translation / QC pipeline
├── results/    # Canonical per-language result CSVs (the numbers in the paper)
└── README.md

The full data — per-language prompts, energy/accuracy/token measurements, and the translated GSM8K & LM-Arena prompts — lives on the 🤗 Hugging Face dataset. Energy is measured with the ML.ENERGY Benchmark (vLLM serving + the Zeus energy-measurement library), reporting steady-state per-token energy.

results/ files

File	Contents
belebele_canonical_Qwen3-8B_0shot.csv	Main experiment: 122 languages × {energy/token, output tokens, total energy, accuracy}
belebele_canonical_{Qwen3-14B,Qwen3-8B,Llama-3.1-8B,Llama-3.1-70B}.csv	Cross-model per-language energy
cross_model_12lang_table.csv	Cross-model comparison table
crosstask_seqs256_summary.csv	Cross-task (Belebele / GSM8K / LM-Arena), 8-language subset
belebele_seqsweep_l40s_0shot.csv	Batch-size sweep, L40S, 8 languages
seqsweep_belebele_v1_qwen8b_RTX6000_all8.csv	Batch-size sweep, RTX 6000 Pro Blackwell, 8 languages

🔬 Experimental setup

• Models: Qwen3-8B / 14B / 32B, gemma-3-27b-it, Llama-3.1-8B-Instruct (cross-model).
• Serving: vLLM, batch size (max_num_seqs) = 256 unless noted; 32B & 27B use pipeline parallelism (degree 2).
• Hardware: NVIDIA L40S (48 GB) and RTX 6000 Pro Blackwell (96 GB).
• Tasks: Belebele (reading comprehension, 122 langs), translated GSM8K (math), LM-Arena (open chat).
• Prompting: zero-shot chain-of-thought; per-language instructions/primers translated with NLLB-200 and quality-controlled by back-translation (BERTScore ≥ 0.85, COMET ≥ 0.75), hand-curated where the automated pipeline failed.

▶️ Reproducing the tables & figures

pip install -r requirements.txt
# Scripts read run outputs from a configurable root (set REPO_ROOT / RUN_ROOT),
# or operate on the CSVs in results/. Examples:
python analysis/build_cross_model_table.py
python analysis/build_crosstask_table_0shot.py
python analysis/build_cross_gpu_table.py
python analysis/build_seqsweep_figs.py

To re-run the energy measurements from scratch, use the ML.ENERGY Benchmark with the models, GPUs, and batch sizes above.

📚 Citation

@article{language-energy-divide,
  title  = {The Language--Energy Divide: Measuring Energy Costs of Multilingual LLM Inference},
  author = {Deng, Naihao and Shen, Alissa and Feng, Yiming and Nwatu, Joan and
            Chung, Jae-Won and Chowdhury, Mosharaf and Chen, Yulong and Mihalcea, Rada},
  year   = {2026}
}

This work builds on the ML.ENERGY Benchmark (Chung et al., NeurIPS 2025 D&B) and the Belebele dataset (Bandarkar et al., ACL 2024).

🙏 Acknowledgements

Energy measurement is powered by Zeus and the ML.ENERGY project. Multilingual data from Belebele; translation via NLLB-200.