---
library_name: vllm
pipeline_tag: text-generation
base_model: Qwen/Qwen3.5-35B-A3B
tags:
  - safetensors
  - qwen3_5_moe
  - moderation
---

# Qwen3.5-35B-A3B Moderation — Sparse (BF16)

Merged LoRA fine-tune of `Qwen/Qwen3.5-35B-A3B` for chat content moderation (sparse output format).

- **Base model**: Qwen/Qwen3.5-35B-A3B
- **Format**: BF16
- **Task**: 5-category chat moderation (underage, bestiality, selfHarm, sexualViolenceGore, realTerrorism)
- **Output**: Sparse JSON — `{}` for safe, `{"underage": "evidence"}` for flagged
- **Serving**: vLLM with `--tensor-parallel-size 1` on 1xH200 or `--tensor-parallel-size 2` on 2xH100, requires CUDA 12.6+