AmeriXLM

Multilingual language model pre-trained on Latin American Spanish and Brazilian Portuguese, with additional coverage of English (Americas) and indigenous languages of the hemisphere.

Model Description

AmeriXLM corrects a systematic gap in existing multilingual models: XLM-R and mBERT allocate training compute proportional to available text volume, which underrepresents Latin American regional variants even within Spanish and Portuguese allocations. AmeriXLM is trained on a corpus that explicitly weights Latin American regional text.

Architecture: RoBERTa
Parameters: 270M
Vocabulary: 64,000 tokens trained on regional corpus
Context length: 512 tokens

Training Data

Brazilian Portuguese: 45B tokens (news, government, legal, web)
Latin American Spanish: 38B tokens (news, government, legal, web)
English Americas: 12B tokens
Indigenous languages: 800M tokens (Quechua, Guarani, Nahuatl)

Benchmark Results

Benchmark	Score	XLM-R large baseline	Delta
AmeriNLI es-LA	83.7	79.4	+4.3
AmeriNLI pt-BR	82.9	78.2	+4.7
RegioNER es-LA F1	88.2	86.7	+1.5
RegioNER pt-BR F1	87.6	85.3	+2.3

Usage

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("genia-americas/ameri-xlm") model = AutoModel.from_pretrained("genia-americas/ameri-xlm")

inputs = tokenizer("Texto en español latinoamericano", return_tensors="pt") outputs = model(**inputs)

Citation

@techreport{genia2025amerixlm, title={AmeriXLM: A Multilingual Language Model for the Americas}, author={GENIA Americas Corporation}, institution={GENIA Americas / RaceFor.AI}, year={2025}, url={https://github.com/GENIA-Americas/multimodal-ai-americas} }