guicybercode/Aira-2-portuguese-124M-mlx-4bit

Aira-2-portuguese-124M MLX 4-bit

This is nicholasKluge/Aira-2-portuguese-124M, a Portuguese language model, converted to 4-bit MLX so it can run with mlx-lm (size in the Files table below). I only tested it on the MLX CPU backend on Linux, not on Apple Silicon.

pip install mlx-lm
mlx_lm.generate --model guicybercode/Aira-2-portuguese-124M-mlx-4bit --prompt "<|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>" --max-tokens 32 --temp 0

All credit for the model goes to the original authors, Nicholas Kluge Corrêa (nicholasKluge). I only converted the format. License: Apache-2.0, same as the original.

Versão quantizada do modelo nicholasKluge/Aira-2-portuguese-124M para uso local; todo o crédito é dos autores originais.

Files

File	Quantization	Size
model.safetensors	4-bit (group size 64), non-quantized tensors in float16	70 MB (70197083 bytes)

How this was made

mlx-lm 0.32.2 with mlx 0.32.4 (CPU backend, Linux x86_64):

mlx_lm.convert --hf-path nicholasKluge/Aira-2-portuguese-124M --mlx-path Aira-2-portuguese-124M-mlx-4bit -q --q-bits 4 --dtype float16

Before converting, the weight names in model.safetensors were renamed from transformer.* to * and the tied lm_head.weight was dropped, because mlx-lm's gpt2 loader expects the original GPT-2 key layout and rejected the file otherwise (Received 148 parameters not in model). No weight values were changed; the converted copy was passed to --hf-path instead of the repo id.

--dtype float16 was used because, on the MLX CPU backend on Linux, a bfloat16 conversion produced only <unk> tokens in testing, while float16 produced normal text. Not tested on Apple Silicon.

Usage

pip install mlx-lm
mlx_lm.generate --model guicybercode/Aira-2-portuguese-124M-mlx-4bit --prompt "<|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>" --max-tokens 32 --temp 0

Smoke test

Prompt: <|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>, 32 tokens, temperature 0, run on the MLX CPU backend on Linux. Only a load-and-generate check, not an evaluation.

A capital do Brasil é Brasília.

No benchmarks were run on this quantized model. See the original model card for the authors' evaluations and intended use.