Large language models have reshaped machine translation by unifying context understanding, terminology control, and stylistic adaptation inside a single inference step. Unlike earlier neural machine translation systems that rely on dedicated encoder-decoder architectures, modern LLMs treat translation as a conditional generation task, leveraging broad pretraining on multilingual corpora to handle low-resource languages, idiomatic phrasing, and domain-specific jargon without retraining the entire model.
Why LLMs for Machine Translation
Traditional NMT systems excel at high-throughput sentence-level translation but struggle with discourse-level coherence, pronoun resolution across paragraphs, and adaptive tone. LLMs mitigate these issues through long-context windows and instruction tuning. A single prompt can encode formatting constraints, glossary terms, and target-audience registers, which would otherwise require separate preprocessing pipelines or fine-tuned adapters.
Additionally, LLMs support zero-shot and few-shot translation. Providing three to five parallel examples inside the prompt often matches or exceeds the quality of domain-adapted NMT models, reducing the engineering overhead of maintaining separate models per language pair or vertical.
Prompt Engineering for Translation
Translation quality depends heavily on prompt structure. A robust pattern separates the source text from instructions and includes constraints such as source and target language tags, domain or style guidelines, glossary constraints, and output format.
For example, a system prompt might read:
You are a professional translator. Translate the user-provided text from Spanish to English. Preserve markdown formatting. Use the provided glossary: "chatbot" maps to "agente conversacional".
For agentic workflows, you can chain translation with post-editing or quality estimation by calling the model multiple times with different system prompts. JSON mode is useful when you need structured output containing translated segments alongside confidence scores or terminology mappings.
Implementation with the OpenAI SDK
Oxlo.ai exposes a fully OpenAI-compatible API, so you can use the official Python SDK by swapping the base URL and model name.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.oxlo.ai/v1",
api_key=os.environ.get("OXLO_API_KEY")
)
def translate_text(source_text, source_lang="German", target_lang="English"):
response = client.chat.completions.create(
model="qwen3-32b", # or llama-3.3-70b, kimi-k2.6
messages=[
{
"role": "system",
"content": (
f"You are an expert translator. Translate from {source_lang} to {target_lang}. "
"Preserve formatting, honorifics, and technical terms. "
"Do not add explanations."
)
},
{
"role": "user",
"content": source_text
}
],
temperature=0.3,
max_tokens=4096
)
return response.choices[0].message.content
document = """
## Vertrag
Der Auftragnehmer verpflichtet sich, die Dienstleistungen fristgerecht zu erbringen...
"""
print(translate_text(document))
This pattern works identically across the Oxlo.ai model catalog. Function calling and streaming responses are also available for real-time translation interfaces.
Model Selection on Oxlo.ai
Oxlo.ai hosts 45+ models across seven categories, several of which are well suited for translation workloads:
- Qwen 3 32B: Strong multilingual reasoning and agent workflows; effective for high-accuracy translation between low-resource language pairs.
- Llama 3.3 70B: General-purpose flagship with broad language coverage; a reliable default for production translation pipelines.
- DeepSeek R1 671B MoE: Deep reasoning capabilities help disambiguate complex legal, medical, or technical passages before translating.
- Kimi K2.6: Advanced reasoning with a 131K context window, ideal for translating full documents while maintaining cross-sentence coherence.
- DeepSeek V4 Flash: Efficient MoE architecture with a 1M context window, enabling near-state-of-the-art reasoning on entire books or lengthy technical manuals in a single request.
- GLM 5: 744B MoE parameter count supports long-horizon agentic tasks, useful when translation is one step in a larger content-processing pipeline.
For latency-sensitive applications, smaller models such as Qwen 3 Coder 30B or Oxlo.ai Coder Fast can handle structured formats such as JSON or code comments with minimal overhead.
Handling Long Documents
Document-level translation introduces coreference resolution, consistent terminology, and formatting preservation across thousands of tokens. Chunking text into sentence-level segments destroys coherence. LLMs with extended context windows solve this by ingesting entire chapters or reports in one pass.
On Oxlo.ai, models such as Kimi K2.6 and DeepSeek V4 Flash support 131K and 1M token contexts respectively. This eliminates the need for fragile chunking logic and reduces pipeline complexity.
Cost Considerations
Token-based pricing penalizes long-context translation. A 50-page legal contract may contain tens of thousands of input tokens, and with per-token billing, each translation pass incurs substantial cost that scales linearly with document length.
Oxlo.ai uses flat per-request pricing: one fixed cost per API call regardless of prompt
Top comments (0)