Supported Languages
V3 Languages
Starting in V3, our models will be trained in groups as opposed to single language pairs.Unsupported Languages
Enlarges the table by opening it in a full screen dialogOpenLarge Language Models
LILT is shipped with:- META: Llama-3-70B-Chat-AWQ
- OpenAI: Whisper-large-v3
GPU Requirements
V4 language models require a minimum of three dedicated GPUs:
vmf-service runs on CPU — after the B25/BM25 retrieval migration, most language pairs no longer require the dense embedding model, so it no longer counts against the GPU budget.rayman is the neural translation service and runs on a single GPU by default. The neural v4.0 pipeline that provides optional Uzbek and Chechen language support requires one additional GPU. If that optional feature is disabled, rayman runs on a single GPU and the baseline total is four GPUs.Each GPU (or GPU split) must provide at least 24GB of VRAM (for example, an NVIDIA L4). This is required to run the translate pod successfully.
Costs
For accurate figures, make use of the AWS cost calculator or refer to your cloud provider’s infrastructure cost tool.New Language Support
We add new AI-supported languages to the LILT Platform continuously based on three criteria:- community and customer demand
- availability of training data
- baseline translation quality that is competitive with Google Translate and Microsoft Translator