# Quantization and reproducibility record

## Source

- Model: `nvidia/Nemotron-3-Embed-1B-BF16`
- Revision: `0677b2025cbc37daf92d7b9c7a225de8cfbc5b0b`
- Source format: BF16 Safetensors
- Upstream license: OpenMDW License Agreement 1.1

## Toolchain

- Project: `ggml-org/llama.cpp`
- Tag/build: `b10015`
- Commit: `12127defda4f41b7679cb2477a4b0d65ee6a0c8f`
- Converter: `convert_hf_to_gguf.py`
- Quantizer: `llama-quantize`
- Importance matrix: none

Equivalent command outline (local paths omitted):

```bash
python convert_hf_to_gguf.py \
  nvidia--Nemotron-3-Embed-1B-BF16-snapshot \
  --outfile nemotron-3-embed-1b-f16.gguf \
  --outtype f16

llama-quantize \
  nemotron-3-embed-1b-f16.gguf \
  nemotron-3-embed-1b-q4_k_m.gguf \
  Q4_K_M
```

## Output identity

- File: `nemotron-3-embed-1b-q4_k_m.gguf`
- Size: `749352096` bytes
- SHA-256: `9a74166f51dbc280073748fa199bea49283bd21f7f9280f2dec2b4d975ddfd1d`
- GGUF version: 3
- Quantization version: 2
- GGUF file type: 15 (`Q4_K_M`)
- Tensor count: 146
- Metadata/data boundary: byte `7869600`

Selected embedded architecture metadata:

| Key | Value |
|---|---:|
| `general.architecture` | `mistral3` |
| `general.name` | `nvidia/Nemotron-3-Embed-1B-BF16` |
| `general.finetune` | `0677b2025cbc37daf92d7b9c7a225de8cfbc5b0b` |
| `mistral3.block_count` | 16 |
| `mistral3.context_length` | 262144 |
| `mistral3.embedding_length` | 2048 |
| `mistral3.feed_forward_length` | 6144 |
| `mistral3.attention.head_count` | 24 |
| `mistral3.attention.head_count_kv` | 8 |
| `mistral3.attention.causal` | false |
| `mistral3.pooling_type` | 1 (mean) |
| `mistral3.vocab_size` | 131072 |

## Integrity check

```bash
shasum -a 256 nemotron-3-embed-1b-q4_k_m.gguf
```

Expected output:

```text
9a74166f51dbc280073748fa199bea49283bd21f7f9280f2dec2b4d975ddfd1d  nemotron-3-embed-1b-q4_k_m.gguf
```