GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
HoshiuZ
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
HoshiuZ
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Evaluation used text-only input on the frozen 3,668-item WMDP C set. This does not evaluate image or broader multimodal safety. The reported C result is from the frozen scoring run; this exported HF checkpoint was separately compared against the same edited full wrapper on CPU, tensor by tensor.
For base accuracy A0 = base_correct / n and edited accuracy Au = edited_correct / n, the project reports C = (A0 - Au) / (A0 - 0.25). The exact result is shown as a reduced fraction as well as a percentage.
| WMDP C items | Base correct | Edited correct | Exact C | C |
|---|---|---|---|---|
| 3668 | 2272 | 1073 | 1199/1355 | 88.4871% |
The separate text-only S evaluation is complete (raw S: 99.199443%; capped S: 99.199443%). See the S results at the frozen project commit. No image-safety or general multimodal-safety claim is made.
Qwen/Qwen3-VL-8B-Instruct0c351dd01ed87e9c1b53cbc748cba10e6187ff3bQwen3VLForConditionalGeneration53b2ed3e8b9b3a21884d48ec9c784de67d6b6b9d4f6e0c912ff4326833f30d68 (402,656,007 bytes)down_proj matrices in layers 4, 5, 6, 7.The sparse edit was produced by a cyber-direction GROM closed-form edit followed by the recorded NPO steps. Only the listed down_proj matrices are included in the update.
Use the complete model through Transformers; no training code or overlay application is needed at load time.
python
import torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGenerationrepo = "HoshiuZ/Qwen3-VL-8B-Unlearned"processor = AutoProcessor.from_pretrained(repo)model = Qwen3VLForConditionalGeneration.from_pretrained( repo, torch_dtype=torch.bfloat16, device_map="auto")The source checkpoint is marked Apache-2.0. See LICENSE, Apache License 2.0, and the source model card. This repository does not claim Qwen affiliation or endorsement.
Research and implementation source: HoshiuZ/multimodal-llm-safety.
import torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGenerationrepo = "HoshiuZ/Qwen3-VL-8B-Unlearned"processor = AutoProcessor.from_pretrained(repo)model = Qwen3VLForConditionalGeneration.from_pretrained( repo, torch_dtype=torch.bfloat16, device_map="auto")