riva-translate-4b-instruct-v1_1

Riva-Translate-4B-Instruct-v1.1 is a 4-billion-parameter instruction-tuned neural machine translation model from NVIDIA, designed to translate natural language text between twelve supported languages. The language set spans English, German, European and LATAM Spanish, French, Brazilian Portuguese, Russian, Simplified and Traditional Chinese, Japanese, Korean, and Arabic, giving the model coverage across major European, East Asian, and Middle Eastern language families as well as both regional variants of Spanish and Chinese. NVIDIA positions this release as a step forward in its in-house NMT lineage, claiming a meaningful performance uplift over earlier Riva translation models, though the supplied evidence is limited to the company's own reference pages rather than independent third-party benchmarks.

The model is built for practical translation workloads in both consumer and enterprise settings, supporting fine-grained sentence-level outputs as well as longer document-level translation passages that require cross-sentence context. Because the weights are released openly under the NVIDIA Community Model License with Apache 2.0 terms noted as additional information, developers and researchers can self-host the model or evaluate it through the NVIDIA NIM trial, with the trial deployment governed by the NVIDIA API Trial Terms of Service. The intended audience includes translators and localization teams working across the supported language pairs, where the combination of broad language coverage, document-level capability, and open-weight availability makes it a flexible choice for building multilingual pipelines, translation assistants, or content localization tools.