Vistral-7B

Vistral-7B is an open-weight Vietnamese language model built by the Viet-Mistral community on the Mistral 7B backbone, with a chat variant (Vistral-7B-Chat) tuned for Vietnamese conversation. It has served as the reference monolingual Vietnamese baseline in SEA-language evaluations — SeaLLMs' own VMLU benchmark reporting notes that Vistral-7B is the monolingual model multilingual systems like SeaLLM-7B-v2 are measured against at similar sizes. A vision-language sibling, Vistral-V-7B, extends the model family to Vietnamese image description via the LLaVA approach with a SigLIP encoder.\n\nThe honest picture: Vistral is a community fine-tune, not a from-scratch pretrain like PhoGPT — it inherits Mistral's strengths and its blind spots in Vietnamese. Documentation is thin compared to lab releases, and the exact release date is not well documented (it predates 2025). Still, as the de-facto open Vietnamese chat baseline, it belongs in any survey of regional models.