Unlike other LLMs that rely on massive scaling, DeepSeek models use a more efficient training approach allowing them to be trained on lower-end hardware while maintaining high efficiency.
As ETH Zurich professor Martin Vechev explained, the implementation of the Mixture of Experts (MoE) technique with key improvements sets these models apart. (MoE allows different model parts (experts) to specialize in tasks, activating only a subset during inference to boost efficiency and potentially improve accuracy.)
Simone Bohnenberger-Rich, Chief Product Officer at Phrase, attributed DeepSeek’s effectiveness to high-quality training data, underscoring the critical role of data quality in AI development.
As with most LLMs, DeepSeek’s models have a wide range of practical applications — including, of course, AI translation.
Users’ Feedback on Translation Capabilities
DeepSeek’s translation capabilities have attracted attention, with users on platforms like X and LinkedIn sharing experiences across several languages.
DeepSeek excels in Chinese-to-English translation, with one user remarking, “We shouldn’t be surprised that DeepSeek does way better” compared to other models in this language pair.
Beyond Chinese, DeepSeek has shown promise in several other languages. Serbian translations were not only accurate but also faster than ChatGPT. In Spanish, it outperformed ChatGPT in a basic translation test, providing both an accurate translation and a follow-up sentence, which allowed the user to stay on the same page without needing to navigate away.
In Turkish, the DeepSeek model was praised for its ability to align closely with a company’s style guide. Czech and Hungarian users said it delivered “awesome quality” — as Claude — across a variety of texts, including legal documents and product translations.
Punjabi speakers reported rapid improvements in translation fluency compared to ChatGPT and Gemini, while Malayalam users appreciated its grammatically accurate and culturally nuanced translations.
When tested on anime subtitles, DeepSeek demonstrated strong contextual understanding, with a user noting that it was “insanely good” at grasping the nuances of translation. Additionally, the translations required minimal revision, with 90% of the sentences flowing naturally from the start.
However, despite the overwhelmingly positive feedback, some users still prefer other models like Claude or Copilot for translation purposes.
For businesses, the adoption of DeepSeek has proven cost-effective as well. “DeepSeek is cutting our translation cost by a factor of 50x compared to Google Translate — so we are now integrating DeepSeek to do FR and DE translations,” an early-stage CEO said.
2024 Slator Pro Guide: Translation AI
The 2024 Slator Pro Guide presents 20 new and impactful ways that LLMs can be used to enhance translation workflows.
Language Industry Experts Remain Skeptical
While many are excited about DeepSeek’s technological advances, language industry experts remain skeptical.
Gert van Assche, CTO at Summa Linguae Technologies, acknowledged the R1 model’s logical reasoning capabilities but found the V3 model’s translation performance “rather disappointing.” Van Assche emphasized the need for more extensive testing before drawing conclusions.
Konstantin Dranch, Language Industry Researcher and Founder at Custom.MT, echoed this, calling for structured evaluations from translators, language service providers (LSPs), and localization teams.
Yasmin Moslem, an NLP researcher, tested the DeepSeek-V3 model for medical translation in French and Portuguese. While it performed well, Llama-3.1 405B slightly outperformed it. Additionally, fine-tuning NLLB 3.3B produced high-quality results, underscoring the effectiveness of specialized machine translation models when sufficient domain-specific data is available.
“We’ve seen astonishing step-change breakthroughs in AI at various points in the past 3 years, so what surprised me most was how many people were surprised to see another.” — Georg Ell, CEO, Phrase
From a business perspective, some leaders remain cautious about its adoption. Georg Ell, CEO at Phrase, acknowledged the model’s potential but warned that large enterprises may hesitate to embrace it due to regulatory concerns, especially regarding data privacy.
DeepSeek faces regulatory scrutiny and allegations of unethical data use. Localization orchestration startup Blackbird.io has declined to integrate it, citing concerns over potential bias in content generation, lack of transparency in data storage and third-party data sharing, excessive data collection, and opaque monitoring of user inputs. However, DeepSeek has gained support from several AI infrastructure and deployment platforms, including Amazon SageMaker AI and Bedrock, Azure AI Foundry, Groq, GitHub, and HuggingFace.
Phrase’s Ell also cautioned that the AI industry should become accustomed to rapid advancements like these, suggesting that new breakthroughs should no longer come as a surprise to industry players. “We’ve seen astonishing step-change breakthroughs in AI at various points in the past 3 years, so what surprised me most was how many people were surprised to see another. We need to get used to it!” he said.
Andre Martins, Head of Research at Unbabel, sees DeepSeek as proof that open models can compete with leading closed systems. He advocates for scientific transparency to accelerate AI progress and warns that Europe must develop its own solutions rather than relying on external technologies.
More broadly, though, DeepSeek’s new launch has created global pushback with Western companies and nation-states trying to restrict access and the United States looking into how a Chinese startup got their hands on what some speculate were large numbers of presumably export-controlled NVIDIA chips.
