inclusionAI released SingProbe, an intrinsic streaming guardrail for Qwen3.5-397B-A17B
inclusionAI released Qwen3.5-397B-A17B-singprobe, an Apache-2.0 licensed intrinsic streaming guardrail probe built on Qwen/Qwen3.5-397B-A17B, on Hugging Face. It adds less than 0.5% decode-time overhead and uses 8.13M parameters tapped at layers [18, 38, 58].
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 probe from Hugging Face and integrate it with SGLang or vLLM to add streaming safety scoring to Qwen3.5-397B-A17B.
- For investors
- The release of an open-weights safety probe by an Ant Group-affiliated entity may indicate a strategy to build ecosystem trust and adoption for Qwen models.
inclusionAI released Qwen3.5-397B-A17B-singprobe on Hugging Face under Apache-2.0. The probe reuses the base model's hidden states during generation 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 adds 8.13M parameters and taps layers [18, 38, 58] of the base model. Reported benchmark results include F1 0.8750 for query intent classification, F1 0.8696 for response safety classification, R-AUC/T-AUC 0.9905/0.9339 for streaming safety, and AUC 0.8117 for hallucination detection. Benign-response false-positive rate is 0.03% average across 5 datasets.
Developers using Qwen3.5-397B-A17B can add streaming safety scoring with less than 0.5% decode overhead, avoiding a separate safety model. This may reduce inference cost and latency for guardrailed deployments compared to external classifiers.
The probe is open-weights under Apache-2.0, allowing commercial use without licensing fees. Its low overhead and integration with popular serving frameworks could lower the cost of adding safety guardrails to Qwen3.5-397B-A17B deployments.
Adoption can be checked by monitoring downloads and community usage of the Hugging Face model page, and by whether SGLang or vLLM integrations are updated in their main branches.