inclusionAI open-sourced Qwen3.5-4B-singprobe, a streaming guardrail probe for Qwen3.5-4B
inclusionAI released Qwen3.5-4B-singprobe, an Apache-2.0 licensed streaming guardrail probe built on Qwen/Qwen3.5-4B, on Hugging Face. It adds less than 0.5% decode-time overhead and scores query intent, response unsafety, and hallucination risk at every token.
China context
- Original name
- inclusionAI
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the Apache-2.0 weights from Hugging Face and integrate the probe via SGLang or vLLM branches.
- For investors
- The release shows Ant Group's inclusionAI is actively contributing open-weight safety tooling, which may influence enterprise adoption of Qwen models.
inclusionAI released Qwen3.5-4B-singprobe on Hugging Face under Apache-2.0. The model is an intrinsic streaming guardrail that reuses the base model's hidden states to score query intent, response unsafety, and hallucination risk at every token, adding less than 0.5% decode-time overhead. It is supported through SGLang and vLLM integration branches.
The probe taps layers [9, 19, 30] of Qwen3.5-4B and outputs 8 intents plus unsafe and hallucination scores. Reported F1 is 0.8650 for query intent classification, 0.8646 for response safety, R-AUC/T-AUC of 0.9870/0.9286 for streaming safety, and AUC 0.7855 for hallucination detection. Benign-response false-positive rate is 0.04% average across 5 datasets.
Developers deploying Qwen3.5-4B can add streaming safety and hallucination checks without running a separate safety model, reducing inference cost and latency. Competitors offering separate guard models face pressure to match this integrated approach.
For enterprises using Qwen3.5-4B, this probe offers a low-overhead way to monitor and mitigate unsafe or hallucinated outputs in real time, potentially reducing compliance and reputational risk.
Next signals to check: adoption of the SGLang/vLLM integration branches, independent benchmarks against the reported metrics, and whether inclusionAI extends SingProbe to other base models.