Custom.MT has integrated three advanced translation models — Widn Vesuvius, Lara, and Claude 3.5 Sonnet — into widely used CAT tools and TMS platforms, including Trados, memoQ, Smartling, and XTM.
The 2024 WMT machine translation benchmarks, organized by leading researchers, highlighted the increasing dominance of large language models (LLMs) over traditional machine translation (MT) systems. LLMs like Unbabel’s Tower (available as Widn) and Claude 3.5 demonstrated superior translation quality, when evaluated by professional linguists on news content.
In addition to quality, LLMs offer cost advantages, being 2–3 times more affordable than fine-tuned engines. They can also be tailored to specific requirements through prompt-based customization, reducing reliance on engineering or computational resources.
Overview of the Models
- Widn Vesuvius: Achieved the highest performance among participants in this year’s WMT translation benchmarks, surpassing systems such as Google Translate and DeepL in evaluations conducted by professional translators.
- Lara: A model developed by Translated, recognized for its fluency and accuracy.
- Claude 3.5 Sonnet: A top scorer at WMT, developed by Anthropic. In combination with DeepL for automated post-editing, it can reduce errors by more than 60%.
The new models are now accessible via Custom.MT Console app (https://console.custom.mt/) for integration into supported platforms. Users are encouraged to evaluate their performance across various content types and provide feedback.
