22 models · ranked by HuggingFace downloads
T5-small is the 60M-parameter variant of Google's Text-to-Text Transfer Transformer, casting all NLP tasks as seq2seq problems. It was influential in establishing the unified text-to-text training paradigm but is outdated for production use.
T5-base is the 220M-parameter version of Google's T5, providing a better accuracy-speed trade-off than T5-small for seq2seq tasks while remaining significantly smaller than T5-large. It established the text-to-text format for multi-task NLP.
nllb-200-distilled-600M is an encoder-decoder translation model. Given a source sentence, it generates the corresponding target-language string.
opus-mt-nl-en translates text between languages using a sequence-to-sequence architecture. It accepts source text and generates a target-language equivalent.
opus-mt-en-de performs neural machine translation. Quality depends on language pair and domain; low-resource pairs and specialized vocabulary benefit from additional fine-tuning.
vntl-llama3-8b-v2-gguf translates text between languages using a sequence-to-sequence architecture. It accepts source text and generates a target-language equivalent.
opus-mt-fr-en performs neural machine translation. Quality depends on language pair and domain; low-resource pairs and specialized vocabulary benefit from additional fine-tuning.
opus-mt-es-en is Helsinki-NLP's MarianMT model for Spanish-to-English machine translation, trained on the OPUS multilingual corpus. It is one of the most widely used open neural MT models for this language pair, offering fast inference via the MarianMT architecture without requiring a large GPU. The model is part of a family covering hundreds of language pairs with a consistent interface.
opus-mt-en-es is a Helsinki-NLP Marian-based neural machine translation model for EN-to-Spanish translation, trained on OPUS parallel corpora. It uses the MarianMT architecture with SentencePiece tokenization and produces competitive translations for common language pairs. The model is available through the Hugging Face translation pipeline without additional dependencies.
opus-mt-tr-en translates text between languages using a sequence-to-sequence architecture. It accepts source text and generates a target-language equivalent.
opus-mt-it-en is a Helsinki-NLP Marian-based neural machine translation model for Italian-to-EN translation, trained on OPUS parallel corpora. It uses the MarianMT architecture with SentencePiece tokenization and produces competitive translations for common language pairs. The model is available through the Hugging Face translation pipeline without additional dependencies.
opus-mt-ar-en is a Helsinki-NLP Marian-based neural machine translation model for Arabic-to-EN translation, trained on OPUS parallel corpora. It uses the MarianMT architecture with SentencePiece tokenization and produces competitive translations for common language pairs. The model is available through the Hugging Face translation pipeline without additional dependencies.
opus-mt-de-en performs neural machine translation. Quality depends on language pair and domain; low-resource pairs and specialized vocabulary benefit from additional fine-tuning.
Part of the Helsinki-NLP OPUS-MT family, this Marian-based model translates English to Modern Standard Arabic. It was trained on OPUS parallel corpora and supports inference via PyTorch, TensorFlow, and Rust backends. With 47 likes and over 400K downloads, it is among the more widely used Helsinki-NLP translation checkpoints.
T5-large (770M) is Google's text-to-text transfer transformer pretrained on C4, framing all NLP tasks as text generation. The large variant provides a significant quality step over T5-base for sequence-to-sequence tasks while remaining practical for fine-tuning.
madlad400-3b-mt is an encoder-decoder translation model. Given a source sentence, it generates the corresponding target-language string.
Built for machine translation, opus-mt-en-it is a model with publicly available weights. opus-mt-en-it is Apache 2.0-licensed, clearing it for closed-source and paid products. Check the opus-mt-en-it model card for benchmarks and intended use before adopting it.
opus-mt-ko-en is an encoder-decoder translation model. Given a source sentence, it generates the corresponding target-language string.
Hunyuan-MT-7B-GGUF is an open-weight machine translation model. At about 7000M parameters, Hunyuan-MT-7B-GGUF sits in the mid-sized tier, which sets its memory and latency budget. Prebuilt GGUF weights make local and edge inference of Hunyuan-MT-7B-GGUF straightforward. Hunyuan-MT-7B-GGUF is community-maintained, so track upstream changes and pin a known-good revision.
Built for machine translation, Sugoi-14B-Ultra-GGUF is a model with publicly available weights. GGUF builds of Sugoi-14B-Ultra-GGUF are published alongside the full checkpoint for low-memory serving. Sugoi-14B-Ultra-GGUF is Apache 2.0-licensed, clearing it for closed-source and paid products. Before relying on Sugoi-14B-Ultra-GGUF, reproduce its key numbers on representative inputs.
t5-3b is an encoder-decoder translation model. Given a source sentence, it generates the corresponding target-language string.
OPUS-MT English-to-French is Helsinki-NLP's Marian NMT model trained on the OPUS multilingual corpus for EN→FR translation. It's a lightweight, fast neural machine translation model using the MarianMT architecture. Apache-2.0 licensed and available in PyTorch, TF, and JAX.