Instructions to use AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 with Transformers:

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

pipe = pipeline("image-text-to-text", model="AI45Research/REX-Gemma-4-26B-A4B-Bio-F3")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("AI45Research/REX-Gemma-4-26B-A4B-Bio-F3")
model = AutoModelForMultimodalLM.from_pretrained("AI45Research/REX-Gemma-4-26B-A4B-Bio-F3", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/AI45Research/REX-Gemma-4-26B-A4B-Bio-F3

SGLang

How to use AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 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 "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3" \
    --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": "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3",
		"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 "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3" \
        --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": "AI45Research/REX-Gemma-4-26B-A4B-Bio-F3",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 with Docker Model Runner:
```
docker model run hf.co/AI45Research/REX-Gemma-4-26B-A4B-Bio-F3
```

REX-Gemma-4-26B-A4B-Bio-F3

Paper | Code | Collection

A REX checkpoint of Gemma-4-26B-A4B with WMDP-Bio knowledge relocated into 8 removable experts per MoE layer (136 experts in total). Excising these experts together with their router rows yields A3, a standard 128-expert model with that knowledge removed.

This repository holds the expanded checkpoint F3: the REX model after 250M tokens of standard full-parameter SFT, with the removable experts still attached. A3 differs from F3 only by excision.

Quick start

git clone https://github.com/AI45Lab/REX.git && cd REX
UV_PROJECT_ENVIRONMENT="$PWD/.venv-cpu" uv sync --project envs/cpu --locked
.venv-cpu/bin/hf download AI45Research/REX-Gemma-4-26B-A4B-Bio-F3 --local-dir ./F3
REX_VENV="$PWD/.venv-cpu" bash scripts/excise_experts.sh ./F3 ./A3

CPU only and about 1 min, with at least 16 GiB RAM. Disk: 52 GiB for F3 plus 49 GiB for A3. ./A3 must not exist beforehand, and expert_expansion_manifest.json must stay in ./F3. Load ./A3 with the model class named in its config.json.

Results

Benchmark	F3 (before excision)	A3 (after excision)
WMDP-Bio (target)	76.43 ± 1.19	41.32 ± 1.38
WMDP-Cyber	50.98 ± 1.12	49.67 ± 1.12
MMLU	73.92 ± 0.34	70.52 ± 0.36
ARC-Challenge (25-shot)	71.84 ± 1.31	67.92 ± 1.36
HellaSwag (10-shot)	85.21 ± 0.35	84.81 ± 0.36
GSM8K (5-shot, strict)	69.22 ± 1.27	68.16 ± 1.28
IFEval (strict prompt)	26.06 ± 1.89	28.10 ± 1.93

Scores in %, ± standard error. lm-eval 0.4.12 with vLLM 0.19.1; IFEval uses the chat template with thinking off. WMDP is 4-way multiple choice (chance = 25%). Comparisons with nine unlearning baselines are in the paper.

Notes

Base model: google/gemma-4-26B-A4B at 24548b62aa02.
Training data is not released. run.json is the redacted training record of F3.
Only text capabilities were evaluated.

License and citation

@misc{jin2026rex,
  title         = {Towards Architectural-Level Safety Control: Relocating Hazardous Knowledge into Removable Experts in Mixture-of-Experts Models},
  author        = {Jin, Cheng and Ma, Zhongtian and Wang, Xiao and Qi, Han and Wen, Xiaoyu and Wang, Xiang and Sun, Youbang and Hu, Xia and Zhou, Bowen and Lu, Chaochao and Zhang, Qiaosheng},
  year          = {2026},
  eprint        = {2610.25791},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2610.25791}
}