A Selective Approach to Translation AI

Huawei is rethinking the use of large language models (LLMs) in translation workflows. In a paper published in May 2025, the company introduced a hybrid AI translation system that determines, before any translation occurs, whether to use a traditional neural machine translation (NMT) engine or a more powerful large language model (LLM).

Unlike earlier approaches that rely on post-translation quality estimation, Huawei’s system analyzes each source sentence in advance. Using a classifier trained on sentence complexity and domain, it predicts whether the LLM would meaningfully outperform NMT. This preemptive decision saves time, energy, and computational cost while preserving translation quality.

When to Call in the Big Models

According to the researchers, the hybrid AI translation system used LLMs for about 25% of sentences across various language pairs: Chinese–English, English–Chinese, Japanese–English, and German–English. This selective application still matched or exceeded full-LLM performance, proving that smarter deployment trumps brute-force usage.

The system defaulted to LLMs for literary or informal content and leaned on NMT for technical or structured language. This not only improves efficiency but also aligns with real-world needs in industries like localization, healthcare, and legal translation.

“If NMT performs equally to LLM,” the authors noted, “we see no improvement after integration.”
The method works best when both engines bring complementary strengths to the table.

Implications for the Translation Industry

This approach could significantly reshape how hybrid AI translation systems are integrated into enterprise and public-sector workflows. Instead of blindly trusting LLMs to handle everything, Huawei shows that strategic use improves both performance and resource allocation.

In a time when businesses seek to balance cost, speed, and quality, selective LLM use offers a scalable and practical path forward.

As generative AI continues to disrupt language technology, solutions that emphasize precision and efficiency over raw power may become the new standard. Huawei’s method sends a clear message: it’s not about using more AI—it’s about using the right AI, at the right time.