DeepL is good. So is Google Translate. So is Microsoft Translator. The neural machine translation revolution of the mid-2010s genuinely improved translation quality, and these engines handle straightforward content remarkably well.

But they all share a fundamental limitation: they translate text. You give them source content, they give you target content, and that’s the entire interaction.

Large language models do something different. They understand context, follow instructions, and adapt their output to requirements. The distinction sounds subtle until you see it in practice.

What MT engines can’t do

Try asking DeepL to translate a marketing headline while keeping it under 50 characters. Or to translate a technical document while using specific terminology from your company glossary. Or to adjust formality level to match your brand voice.

You can’t. The interface is source text in, target text out. Any additional requirements—length constraints, terminology, style, context—have to be handled elsewhere in your workflow.

This is why post-editing exists. MT produces output based purely on the source text. Humans then modify that output to meet all the other requirements the MT engine couldn’t address.

LLMs work differently. You can provide instructions alongside the source text:

• “Translate this marketing copy to German, keeping each line under 60 characters for the UI layout.”
• “Translate this technical document to French, using the following terminology: [glossary].”
• “Translate this customer email to Spanish with a friendly, informal tone appropriate for a B2C brand.”

The model considers these instructions as part of the translation process, not as an afterthought.

Context windows and document awareness

Traditional MT translates segment by segment. Sentence 15 is translated without awareness of sentences 1-14 or 16-30. This causes several problems:

Pronoun resolution. When a sentence contains “it” or “they,” the MT engine doesn’t know what those pronouns refer to. Languages with grammatical gender need to know whether “it” refers to something masculine or feminine.

Terminology consistency. If a technical term appears in sentence 5 and again in sentence 20, MT engines translate each occurrence independently. They might use different target terms for the same source term.

Register drift. The appropriate formality level for a document depends on the whole document, not individual sentences. MT engines can produce jarring register shifts within a single text.

LLMs with sufficient context windows can see entire documents—or at least large portions of them. They maintain consistency across the text because they’re aware of what came before and after each segment.

The instruction-following advantage

The practical impact of instruction-following becomes clear in specific scenarios:

Character-limited UI text. Mobile apps and interfaces have strict length constraints. MT output that’s 20% longer than the source doesn’t fit. With LLMs, you specify the constraint: “Translate to fit within the same character count as the source.” The model adapts its word choices accordingly.

Brand voice and style. Companies invest in brand guidelines that specify tone, vocabulary preferences, and writing style. Feeding these guidelines to an LLM produces translations that match the brand voice. MT produces generic output that post-editors must then adapt.

Terminology enforcement. Professional translation requires consistent use of approved terminology. LLMs can accept glossaries as part of the prompt and apply them during translation, not as a separate lookup step.

Format preservation. When translating structured content—JSON, XML, Markdown—LLMs can be instructed to preserve the structure exactly while translating only the text content. MT engines often struggle with mixed format/content documents.

Research and reference integration

Another capability unique to LLMs is the ability to incorporate reference material into the translation process.

Imagine translating a scientific paper. The translation would benefit from awareness of how other papers in the field have translated key concepts. Or translating a legal document where specific phrasings have established translations in target-language case law.

LLMs can accept reference materials alongside the source text:

• “Translate this patent application to German. Here are relevant German patents for terminology reference.”
• “Translate this software documentation. Here is the translated UI string file to ensure consistency with the product.”
• “Translate this research summary. Here are three published papers on the same topic in the target language.”

The model uses these references to inform translation choices. It’s not just translating words—it’s producing output that fits within an existing body of work.

Quality characteristics: MT vs LLM

Independent evaluations consistently show different quality profiles for MT and LLM translation:

MT strengths:

• Consistent performance on well-represented language pairs
• Predictable output quality
• Fast processing
• Lower cost per word

LLM strengths:

• Better handling of context and nuance
• Superior performance on complex sentences
• More natural-sounding output for creative content
• Ability to follow specific instructions

Common weaknesses:

• Both can make errors that human translators wouldn’t
• Both perform worse on low-resource language pairs
• Both require quality assurance processes

The implication isn’t that LLMs are universally better. It’s that they’re better for different things. High-volume, straightforward content may translate well with MT. Complex, nuanced, or specifically constrained content benefits from LLM capabilities.

The hybrid approach

The most effective translation workflows don’t choose MT or LLM exclusively. They use both strategically:

Stage 1: MT for first pass. Run content through a traditional MT engine for baseline translation. This is fast and inexpensive.

Stage 2: LLM enhancement. Use an LLM to refine the MT output, applying style guidelines, checking terminology, and smoothing unnatural phrasing.

Stage 3: Quality scoring. Evaluate the enhanced translation against quality criteria, flagging segments that need human review.

This approach captures MT’s efficiency while adding LLM’s contextual intelligence. The result is often better than either approach alone, at a cost between pure MT and pure LLM.

Cost considerations

LLM translation is more expensive per word than MT. Frontier model processing costs roughly 10-20x what Google Translate or DeepL costs for equivalent volume—though local model deployment can reduce ongoing costs while ensuring complete data privacy.

Whether this cost difference matters depends on your content:

• If post-editing cost exceeds the LLM premium, LLM translation saves money
• If quality requirements exceed what MT can deliver, LLM translation is necessary regardless of cost
• If content volume is high and quality requirements are modest, MT remains more economical

Most organizations have content in all three categories. The right approach is matching translation methods to content types, not applying one method universally.

The practical shift

For localization teams, LLM translation doesn’t replace existing workflows overnight. It adds capabilities that weren’t previously available at any cost:

• True context-aware translation at scale
• Instruction-following that handles specific requirements automatically
• Reference integration for specialized domains
• Style and brand voice preservation without manual post-editing

These capabilities matter most for content where quality directly impacts business outcomes: customer-facing marketing, product interfaces, legal and regulatory content, brand communications. For that content, the question isn’t whether LLM translation costs more than MT—it’s whether the quality improvement justifies the investment.

For most organizations, for their most important content, it does.

Language Ops supports translation through frontier AI models and traditional MT engines, with local model deployment available for complete data privacy and GDPR compliance. Configurable hybrid workflows combine the strengths of each approach. See the difference on your own content.