GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AI45Research
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AI45Research
Model Tree
Qwen/Qwen3.5-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
bash
git clone --branch v0.1.0 https://github.com/AI45Lab/REX.git && cd REXUV_PROJECT_ENVIRONMENT="$PWD/.venv-cpu" uv sync --project envs/cpu --locked.venv-cpu/bin/hf download AI45Research/REX-Intern-S2-Bio-F3 --revision v0.1.0 --local-dir ./F3REX_VENV="$PWD/.venv-cpu" bash scripts/excise_experts.sh ./F3 ./A3CPU only and about 1 min, with at least 16 GiB RAM. Disk: 67 GiB for F3 plus 65 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.
Intern-S2 ships a custom tokenizer; inspect it before enabling trust_remote_code.
| Benchmark | F3 (before excision) | A3 (after excision) |
|---|---|---|
| WMDP-Bio (target) | 78.63 ± 1.15 | 54.99 ± 1.39 |
| WMDP-Cyber | 58.98 ± 1.10 | 59.49 ± 1.10 |
| MMLU | 84.08 ± 0.30 | 82.55 ± 0.31 |
| ARC-Challenge (25-shot) | 72.01 ± 1.31 | 71.33 ± 1.32 |
| HellaSwag (10-shot) | 85.27 ± 0.35 | 85.12 ± 0.36 |
| GSM8K (5-shot, strict) | 85.22 ± 0.98 | 86.20 ± 0.95 |
| IFEval (strict prompt) | 78.56 ± 1.77 | 78.56 ± 1.77 |
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%).
run.json is a redacted training record.Apache-2.0, Copyright 2026 Shanghai AI Laboratory. Upstream model and tokenizer
licenses also apply; see NOTICE.
bibtex
@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}, howpublished = {\url{https://github.com/AI45Lab/REX}}}git clone --branch v0.1.0 https://github.com/AI45Lab/REX.git && cd REXUV_PROJECT_ENVIRONMENT="$PWD/.venv-cpu" uv sync --project envs/cpu --locked.venv-cpu/bin/hf download AI45Research/REX-Intern-S2-Bio-F3 --revision v0.1.0 --local-dir ./F3REX_VENV="$PWD/.venv-cpu" bash scripts/excise_experts.sh ./F3 ./A3@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}, howpublished = {\url{https://github.com/AI45Lab/REX}}}