How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf guicybercode/bee-350m-pt-base-GGUF:
# Run inference directly in the terminal:
llama cli -hf guicybercode/bee-350m-pt-base-GGUF:
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf guicybercode/bee-350m-pt-base-GGUF:
# Run inference directly in the terminal:
llama cli -hf guicybercode/bee-350m-pt-base-GGUF:
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf guicybercode/bee-350m-pt-base-GGUF:
# Run inference directly in the terminal:
./build/bin/llama cli -hf guicybercode/bee-350m-pt-base-GGUF:
Use Docker
docker model run hf.co/guicybercode/bee-350m-pt-base-GGUF:
Quick Links

cover

bee-350m-pt-base-GGUF

GGUF quantizations of BrCamp/bee-350m-pt-base, a 345.4M-parameter language model pre-trained from scratch in Portuguese by BrCamp. Ready to run on CPU with llama.cpp, Ollama, LM Studio and other GGUF runtimes.

Quantizacoes GGUF do BrCamp/bee-350m-pt-base, modelo de 345,4M parametros pre-treinado do zero em portugues pela BrCamp. Roda em CPU com llama.cpp, Ollama, LM Studio e outros runtimes GGUF.

This is a BASE model: it continues text, it does not follow instructions or chat. / Este e um modelo BASE: ele continua texto, nao segue instrucoes nem conversa.

Usage / Uso

llama.cpp:

llama-cli -hf guicybercode/bee-350m-pt-base-GGUF:Q8_0 -p "O Brasil é um país" -n 64 -no-cnv

Ollama:

ollama run hf.co/guicybercode/bee-350m-pt-base-GGUF:Q8_0

Replace Q8_0 with Q4_K_M or F16 to pick another file. / Troque Q8_0 por Q4_K_M ou F16 para escolher outro arquivo.

Files / Arquivos

File Type Size
bee-350m-pt-base-F16.gguf F16 (no quantization) 692 MB
bee-350m-pt-base-Q8_0.gguf Q8_0 (recommended) 368 MB
bee-350m-pt-base-Q4_K_M.gguf Q4_K_M 253 MB

Note: the hidden size (960) is not a multiple of 256, so in Q4_K_M 192 of 354 tensors use fallback quant types (e.g. q5_0). For this small model Q8_0 is the recommended quality/size trade-off.

Nota: a dimensao oculta (960) nao e multipla de 256, entao no Q4_K_M 192 de 354 tensores usam tipos de fallback (ex.: q5_0). Para este modelo pequeno, Q8_0 e o melhor equilibrio entre qualidade e tamanho.

Sample output / Exemplo (Q4_K_M, greedy, CPU)

Prompt: O Brasil é um país

O Brasil é um país que tem uma economia muito forte, mas que também tem uma grande desigualdade social. A desigualdade social é muito grande, e a desigualdade de renda é muito grande.

Conversion notes / Notas de conversao

  • Converted with llama.cpp convert_hf_to_gguf.py (architecture qwen3) and quantized with llama-quantize.
  • The custom 32k ByteLevel BPE tokenizer is not in llama.cpp's pre-tokenizer list; it was mapped to the gpt-2 pre-tokenizer, which uses the same ByteLevel regex. The original tokenizer also applies NFC normalization, which llama.cpp does not; use NFC-normalized input text for identical tokenization.
  • Context length: 2048 tokens.

Credits / Creditos

All credit for the model goes to BrCamp, authors of bee-350m-pt-base. See the original card for training details, measured results and limitations. License: Apache-2.0, same as the original.

Todo o credito do modelo e da BrCamp. Veja o card original para detalhes de treino, resultados e limitacoes.

Downloads last month
119
GGUF
Model size
0.3B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for guicybercode/bee-350m-pt-base-GGUF

Quantized
(1)
this model