guicybercode/Aira-2-portuguese-124M-GGUF

Aira-2-portuguese-124M GGUF

I wanted to run nicholasKluge/Aira-2-portuguese-124M, a Portuguese language model, on an ordinary CPU without Python or a GPU. This repo has the model converted to GGUF (Q4_K_M, Q8_0 and F16) so it runs in llama.cpp and Ollama.

llama-cli -hf guicybercode/Aira-2-portuguese-124M-GGUF:Q4_K_M -p "<|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>" -n 48

The line above uses llama.cpp's -hf download (tested 2026-10-22 with a llama.cpp build from source).

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	Type	Size
Aira-2-portuguese-124M-Q4_K_M.gguf	Q4_K_M	91 MB (91338144 bytes)
Aira-2-portuguese-124M-Q8_0.gguf	Q8_0	137 MB (136767328 bytes)
Aira-2-portuguese-124M-F16.gguf	F16	253 MB (252581472 bytes)

How these were made

llama.cpp (master, built on Linux x86_64 CPU):

python conv_gpt2pre.py Aira-2-portuguese-124M --outtype f16 --outfile Aira-2-portuguese-124M/Aira-2-portuguese-124M-F16.gguf
llama-quantize Aira-2-portuguese-124M-F16.gguf Aira-2-portuguese-124M-Q4_K_M.gguf Q4_K_M
llama-quantize Aira-2-portuguese-124M-F16.gguf Aira-2-portuguese-124M-Q8_0.gguf Q8_0

The stock converter stopped with "BPE pre-tokenizer was not recognized" because the model's added tokens change the tokenizer hash it checks. The model uses the standard GPT-2 tokenizer (GPT2Tokenizer, vocab.json + merges.txt), so conv_gpt2pre.py is a small wrapper that runs the same convert_hf_to_gguf.py with the pre-tokenizer set to gpt-2. It is included in this repo.

Usage

llama-completion -m Aira-2-portuguese-124M-Q4_K_M.gguf -p "<|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>" -n 48 --temp 0

Smoke test

Prompt: <|startofinstruction|>Qual é a capital do Brasil?<|endofinstruction|>, 48 tokens, temperature 0, raw completion (no chat template). This only checks that the files load and produce Portuguese text; it is not an evaluation.

Q4_K_M:

Qual é a capital do Brasil?A capital do Brasil é Brasília.

Q8_0:

Qual é a capital do Brasil?A capital do Brasil é Brasília.

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