# OpenMLE post-trained checkpoints

Both post-trained checkpoints are publicly downloadable from this repository:

| Model | Path | Weight files |
| --- | --- | ---: |
| OpenMLE-30B | `models/OpenMLE-30B/` | 12 safetensors shards |
| OpenMLE-35B | `models/OpenMLE-35B/` | 15 language-model shards and `model-vision-mtp.safetensors` |

Each directory includes the weight index, configuration, tokenizer and license
files required to load that checkpoint. The 35B vision encoder and MTP
components are inherited from its upstream base; the language-model weights
were post-trained. Model weights use CC BY-NC 4.0; upstream components retain
their Apache 2.0 terms and notices.

This is a **dataset repository** containing model files in subdirectories.
Download only the desired checkpoint, then load its local directory:

```python
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="nono314/openmle",
    repo_type="dataset",
    allow_patterns=["models/OpenMLE-30B/*"],
    local_dir="OpenMLE-release",
)
model_path = f"{root}/models/OpenMLE-30B"
```

For the 35B checkpoint, replace `OpenMLE-30B` with `OpenMLE-35B`.
The respective Transformers architectures are `Qwen3MoeForCausalLM` and
`Qwen3_5MoeForConditionalGeneration`; use a Transformers version supporting
the selected architecture. Do not pass the dataset repository ID directly to
`from_pretrained`; pass the downloaded model directory.

`models/checksums.sha256` lists SHA-256 values for the model files. The
supplementary ZIP contains code and data examples; these large weight files
are hosted here and are not embedded in that ZIP.