MzansiLM is a 125M-parameter decoder-only language model trained from scratch on MzansiText, a multilingual corpus covering all eleven official South African languages.

Paper

Model Details

Parameters: 125,008,384
Architecture: decoder-only LlamaForCausalLM
Hidden size: 512
Intermediate size: 1536
Layers: 30
Attention heads: 9
Key/value heads: 3
Context length: 2048
RoPE theta: 10000.0
RMSNorm epsilon: 1e-5
Tied word embeddings: true
Training attention implementation: flash_attention_2

Tokenizer

MzansiLM uses a custom BPE tokenizer with a vocabulary size of 65536.

[BOS] = 0
[EOS] = 1
[PAD] = 2
[UNK] = 3
Normalizer: NFD
Pre-tokenizer: ByteLevel
Post-processing:
- single sequence: [BOS] $A [EOS]
- pair sequence: [BOS] $A [EOS] [BOS] $B [EOS]

Training Data

The model was trained on MzansiText and covers all eleven official South African languages:

af, en, nso, sot, ssw, tsn, tso, ven, xho, zul, nbl

Related releases:

Paper: arXiv:2603.20732
Raw corpus: anrilombard/mzansi-text
Tokenized corpus: anrilombard/mzansi-text-tokenized
GitHub code and configs: https://github.com/Anri-Lombard/sallm

Intended Use

MzansiLM is a research model for pretraining, fine-tuning, and evaluation on South African languages. It is intended as a reproducible baseline for language modeling and downstream task adaptation.

Usage

MzansiLM uses an explicit attention head dimension (head_dim=56) with hidden_size=512 and num_attention_heads=9. Use Transformers 4.x for this release:

pip install "transformers>=4.52.4,<5" torch accelerate safetensors

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "anrilombard/mzansilm-125m"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

inputs = tokenizer("Namuhla sifunda ngolimi ngoba abantu", return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    pad_token_id=tokenizer.pad_token_id,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

This is a base language model, not an instruction-tuned chat model, so outputs are plain text continuations of the prompt.

Compatibility note

Transformers 5 currently rejects this Llama configuration during validation because 512 is not divisible by 9, even though the model uses the explicit head_dim=56 stored in config.json. Until that validation path supports this configuration, use Transformers 4.x as shown above.

Citation

Please cite the paper:

bibtex

@misc{lombard2026mzansitextmzansilmopencorpus,
      title={MzansiText and MzansiLM: An Open Corpus and Decoder-Only Language Model for South African Languages}, 
      author={Anri Lombard and Simbarashe Mawere and Temi Aina and Ethan Wolff and Sbonelo Gumede and Elan Novick and Francois Meyer and Jan Buys},
      year={2026},
      eprint={2603.20732},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.20732}, 
}

License

Apache License 2.0

uctnlp/mzansilm-125m

MzansiLM 125M