inclusionAI released Qwen3.8-27B-singprobe, a streaming guardrail probe for Qwen3.8-27B
inclusionAI released Qwen3.8-27B-singprobe, an Apache-2.0 licensed streaming guardrail probe built on Qwen/Qwen3.8-27B, 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 model weights from Hugging Face and integrate the probe via SGLang or vLLM branches, enabling streaming safety checks without a separate model.
- For investors
- The release of a low-overhead safety probe by an Ant Group-affiliated entity may indicate a strategic focus on enterprise AI safety, potentially affecting the competitive landscape for AI safety startups.
inclusionAI released Qwen3.8-27B-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 uses 10.1M probe parameters tapped at layers 20, 41, and 62, and outputs 8 intents plus unsafe and hallucination scores. Evaluation results show F1 of 0.8683 on query intent classification, 0.8676 on response safety classification, R-AUC/T-AUC of 0.9881/0.9305 on streaming safety, and AUC of 0.8035 on hallucination detection. The model is supported through SGLang and vLLM integration branches.
SingProbe is a lightweight probe (10.1M parameters) that taps hidden states from layers 20, 41, and 62 of Qwen3.8-27B to produce per-token scores for 8 intents, unsafety, and hallucination. It achieves a benign-response false-positive rate of 0.03% and less than 0.5% decode overhead. The approach avoids running a separate safety model, reducing latency and compute costs for streaming guardrails.
Developers using Qwen3.8-27B can add streaming safety and hallucination detection with minimal overhead, reducing the need for separate guardrail models and lowering inference costs. This may pressure providers of standalone guardrail models to offer tighter integration or lower prices.
The model offers a cost-effective safety solution for enterprises deploying Qwen3.8-27B, potentially reducing compliance overhead and latency. Its Apache-2.0 license allows commercial use, which may accelerate adoption in production systems.
Adoption can be tracked by monitoring downloads and community feedback on the Hugging Face model page, as well as pull requests or issues in the SGLang and vLLM integration branches. Independent benchmarks comparing SingProbe against reference baselines on the same datasets would validate the reported metrics.