inclusionAI released SingProbe, a streaming guardrail probe for Qwen3.8-Flash-Next
inclusionAI released SingProbe, an intrinsic streaming guardrail probe built on Qwen/Qwen3.8-Flash-Next, adding less than 0.5% decode-time overhead. It scores query intent, response unsafety, and hallucination risk at every token. The model is available on Hugging Face under an 'other' license.
China context
- Original name
- inclusionAI
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can integrate the probe via SGLang or vLLM branches to get per-token safety scores with minimal overhead.
- For investors
- Investors can track adoption of this probe as a signal for demand in lightweight safety solutions for open-weight models.
inclusionAI released SingProbe, a lightweight streaming guardrail probe that reuses the base model's hidden states to score query intent, response unsafety, and hallucination risk at every token. It adds less than 0.5% decode-time overhead and is available on Hugging Face under an 'other' license.
The probe taps layers [15, 31, 47] of Qwen3.8-Flash-Next and uses 5.18M parameters to output 8 intents plus unsafe and hallucination scores. Reported F1 scores are 0.8728 for query intent classification, 0.8714 for response safety, and AUC 0.8063 for hallucination detection, with a benign-response false-positive rate of 0.03%.
Developers deploying Qwen3.8-Flash-Next can add streaming safety and hallucination scoring with minimal latency impact, avoiding a separate safety model. This reduces integration cost and complexity for real-time guardrails.
The probe offers a low-overhead method for real-time safety and hallucination monitoring, potentially reducing the need for separate safety models and lowering operational costs for AI deployments.
Adoption can be checked by monitoring downloads and community feedback on the Hugging Face model page, and by whether SGLang or vLLM integrations gain traction.