Instructions to use Jared1728/GPT-OSS-20B-Exact-BF16-Archive with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Jared1728/GPT-OSS-20B-Exact-BF16-Archive with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Jared1728/GPT-OSS-20B-Exact-BF16-Archive")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Jared1728/GPT-OSS-20B-Exact-BF16-Archive", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Jared1728/GPT-OSS-20B-Exact-BF16-Archive with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Jared1728/GPT-OSS-20B-Exact-BF16-Archive"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Jared1728/GPT-OSS-20B-Exact-BF16-Archive",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Jared1728/GPT-OSS-20B-Exact-BF16-Archive

SGLang

How to use Jared1728/GPT-OSS-20B-Exact-BF16-Archive with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Jared1728/GPT-OSS-20B-Exact-BF16-Archive" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Jared1728/GPT-OSS-20B-Exact-BF16-Archive",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Jared1728/GPT-OSS-20B-Exact-BF16-Archive" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Jared1728/GPT-OSS-20B-Exact-BF16-Archive",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Jared1728/GPT-OSS-20B-Exact-BF16-Archive with Docker Model Runner:
```
docker model run hf.co/Jared1728/GPT-OSS-20B-Exact-BF16-Archive
```

GPT-OSS 20B exact BF16 archive backed by MXFP4

This repository contains a lossless archival representation of an expanded BF16 deployment derived from openai/gpt-oss-20b. It stores the expert matrices in their latent MXFP4 form and reconstructs the expanded BF16 checkpoint exactly.

This is an archive, not a drop-in Transformers checkpoint. Decode it first if you want the complete expanded-BF16 directory.

Result

measurement	bytes
expanded BF16 deployment	41,857,491,374
exact GOSBFI archive	11,978,397,252
reduction	71.3829%

Archive SHA-256:

19f11903530ede3ac72b272d64f47ba41b6ff0d6b9327d09e32305f1203d27fb

All 17 deployment files and all nine Safetensors shards were reconstructed byte-for-byte. The decoder verifies every segment and complete output file before promoting it from a temporary file.

Decode

Requirements:

Python 3.10+
NumPy
zstandard
approximately 42 GB of free output space
approximately 5.3 GB peak system RAM in the standalone decoder validation

python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
sha256sum -c CHECKSUMS.sha256
python decode_gosbfi.py \
  gpt-oss-20b-bf16-inverse.gosbfi \
  gpt-oss-20b-bf16 \
  --report decode-report.json
cd gpt-oss-20b-bf16
sha256sum -c ../SOURCE_CHECKSUMS.sha256

The standalone public decoder completed the full validation in 427.96 seconds and reached 5,236,150,272 bytes peak RSS on the development machine. Results vary with storage and CPU speed.

What is public and what is not

This repository includes the minimal decoder and integrity manifests needed to recover and verify the checkpoint. It does not include the encoder, discovery and analysis pipeline, codec-selection experiments, inference integration, profilers, or benchmarking harness.

The decoder necessarily documents the reversible portion of the format. The research process and production tooling are separate from that compatibility surface.

Precision and inference validation

Stored expanded expert BF16 values reconstruct exactly: 19,110,297,600 of 19,110,297,600 BF16 bit patterns matched.
The archive-compatible compressed runtime matched official MXFP4 logits and a deterministic 128-token continuation exactly in the tested path.
Against full expanded-BF16 arithmetic, MXFP4/compressed-runtime logits had 0.4152% relative RMSE and 0.99999166 cosine similarity, with the same top prediction and exact tested eight-token greedy continuation.

Stored-weight exactness and arithmetic-kernel equivalence are distinct claims. The archive round trip is byte-exact. MXFP4 and expanded-BF16 execution are numerically very close but are not claimed to be bit-identical arithmetic.

Licensing, scope, and attribution

This repository uses a file-level license split:

The GOSBFI archive and material derived from the base model remain under the upstream Apache License 2.0. See LICENSE-MODEL-APACHE-2.0.
decode_gosbfi.py is licensed under the PolyForm Noncommercial License 1.0.0. It may be used for noncommercial research, education, experimentation, and evaluation. See LICENSE-CODE-POLYFORM-NONCOMMERCIAL-1.0.0.
Commercial rights to the decoder are not granted by the research license and require a separate agreement with the decoder copyright holder.

See LICENSES.md for the authoritative file mapping.

The base model is OpenAI's openai/gpt-oss-20b, licensed under Apache 2.0. This repository is an independently produced archival packaging and is not an official OpenAI release.

No warranty is provided. Review the upstream model card and license before deployment.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for Jared1728/GPT-OSS-20B-Exact-BF16-Archive

Base model

openai/gpt-oss-20b

Finetuned

(539)

this model