# Local model package

This directory is intentionally excluded from the source repository.

## Layout

- `hf/Confucius4-T3PO/`: copied Hugging Face checkpoint.
- `gguf/Confucius4-T3PO-F16.gguf`: F16 GGUF converted from that checkpoint.
- `gguf/Confucius4-T3PO-Q6_K.gguf`: Q6_K quantization of the F16 GGUF.
- `gguf/Confucius4-T3PO-Q5_K_M.gguf`: Q5_K_M quantization of the F16 GGUF.

## Conversion

- Converter: `ggml-org/llama.cpp`
- Converter commit: `5202104b59ada9005db079eea43882a2b7bf5802`
- Architecture detected by the converter: `Qwen2ForCausalLM`
- GGUF architecture: `qwen2`
- GGUF version: `3`
- Tensor count: `579`

| File | Output type | Size (bytes) |
| --- | :---: | ---: |
| `Confucius4-T3PO-F16.gguf` | `F16` | 29,547,714,080 |
| `Confucius4-T3PO-Q6_K.gguf` | `Q6_K` | 12,124,681,760 |
| `Confucius4-T3PO-Q5_K_M.gguf` | `Q5_K_M` | 10,508,871,200 |

Equivalent local conversion command:

```bash
python3 convert_hf_to_gguf.py \
  model/hf/Confucius4-T3PO \
  --outfile model/gguf/Confucius4-T3PO-F16.gguf \
  --outtype f16 \
  --model-name Confucius4-T3PO
```

The low-bit variants are quantized from the F16 GGUF:

```bash
llama-quantize model/gguf/Confucius4-T3PO-F16.gguf \
  model/gguf/Confucius4-T3PO-Q6_K.gguf Q6_K

llama-quantize model/gguf/Confucius4-T3PO-F16.gguf \
  model/gguf/Confucius4-T3PO-Q5_K_M.gguf Q5_K_M
```

No GGUF splitting was applied.