morena-1.5b-instruct-mlx-4bit

morena-1.5b-instruct (MLX, 4-bit)

An MLX conversion of vamboai/morena-1.5b-instruct, made by Vambo AI. All credit for the model belongs to its author; this repository only changes the file format so it runs on Apple silicon (Macs and iPhones).

• Original model: vamboai/morena-1.5b-instruct, 1.5B parameters, chat, translation and tool calling.
• Author: Isheanesu Misi, Vambo AI. Trained on CINECA Leonardo with support from the AI Hub for Sustainable Development.
• Paper: MORENA: An African Foundation Model (Vambo AI, September 2026).
• Licence: Apache 2.0, the same as the original. See the original repository for the full terms.

Please read the original model card for the evaluation results, intended use, safety measurements and known limitations. They apply unchanged to this conversion, apart from the small quality loss that 4-bit quantization introduces.

Languages

Shona, Swahili, Hausa, Yoruba, Igbo, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans, isiNdebele and Nigerian Pidgin, plus English.

What was changed

1. The original checkpoint is a plain PyTorch decoder (RMSNorm, RoPE, grouped-query attention, SwiGLU, tied embeddings) rather than a transformers architecture. Its weights were renamed one-for-one into the Llama layout (for example layers.N.attn.wq became model.layers.N.self_attn.q_proj) and a matching Llama config.json was written. No weights were retrained, merged or pruned.
2. The weights were quantized with mlx_lm.convert to 4 bits, group size 64 (835 MB of weights).
3. tokenizer.json is the original file, unchanged.

The chat template in tokenizer_config.json reproduces the original format: the prompt starts with <eos>, a user turn opens with the single token <reserved_0> and an assistant turn with <reserved_1>. Do not replace these with <|user|>-style strings; the original authors note that this silently produces degenerate output.

The conversion script is tools/convert_morena.py in the Learn Local project.

Use with mlx-lm

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("byronrode/morena-1.5b-instruct-mlx-4bit")
messages = [{"role": "user", "content": "Translate to isiZulu: I see a cup."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=40))

Citation

Please cite the original work:

@techreport{misi2026morena,
  title  = {MORENA: An African Foundation Model},
  author = {Misi, Isheanesu},
  institution = {Vambo AI},
  year   = {2026},
  month  = {September},
  note   = {Released 18 September 2026}
}