YuCollection/gpt-oss-120b-original-mxfp4-mirror

GPT-OSS-120B — Original MXFP4 Reference Weights (Mirror)

⚠️ IMPORTANT

This is an archived mirror and NOT the original upstream repository maintained by OpenAI.

It is not affiliated with, endorsed by, or officially supported by OpenAI.

It contains only the original reference checkpoints of GPT-OSS-120B in MXFP4 format,
preserved for archival, research, and conversion purposes.

These weights are NOT inference-ready and cannot be used directly for generation without conversion to an appropriate runtime format.
See upstream for inference-ready usage.

---

Repository Scope

This repository intentionally stores a minimal and authoritative subset of files required to preserve the GPT-OSS-120B model.

The repository does not include inference binaries, converted weights, or example usage code.

---

Contents of This Repository

📦 Model Weights (Original Checkpoint)

Located in the original/ directory:

• *.safetensors
  → Original MXFP4 reference weights (sharded)
• model.safetensors.index.json → Tensor-to-shard mapping

These files represent the source-of-truth model parameters released by OpenAI.

📄 Model Configuration & Metadata

• config.json
  → Model architecture definition (layers, dimensions, MoE structure)
• Dtype / quantization metadata
  → Documents the MXFP4 numeric format used by the weights
• LICENSE
  → Apache License, Version 2.0

---

❌ Not Included

To avoid ambiguity, this repository does NOT include:

• Tokenizer files (tokenizer.json, tokenizer.model, etc.)
• Inference-ready weights (Transformers / vLLM / Metal)
• Runtime binaries or compiled artifacts
• Example inference code

Tokenizer and inference artifacts are available from the upstream OpenAI repositories.

---

About MXFP4

The GPT-OSS models were post-trained using MXFP4 quantization of the MoE weights.

• MXFP4 is required to fit GPT-OSS-120B within an 80GB GPU
• Specialized kernels and tooling are required for inference
• This repository preserves the original MXFP4 semantics for future compatibility

---

Intended Use Cases

This repository is suitable for:

• Long-term archival and backup
• Research into model parameters
• Converting weights into runtime-specific formats
• Fine-tuning workflows that support MXFP4
• Audit and reproducibility work

It is not intended for direct inference or deployment.

---

Upstream References

• Official model repository: https://huggingface.co/openai/gpt-oss-120b
• Source code and tooling: https://github.com/openai/gpt-oss
• Model card (arXiv): https://arxiv.org/abs/2508.10925

---

License

This repository is distributed under the Apache License 2.0.

All rights to the model architecture, training, and original release remain with OpenAI.