inclusionAI released Qwen3.5-0.8B-singprobe, a streaming safety probe for Qwen3.5-0.8B
inclusionAI released Qwen3.5-0.8B-singprobe, an Apache-2.0 licensed streaming guardrail probe built on Qwen/Qwen3.5-0.8B, 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 of a lightweight safety probe by an Ant Group-affiliated entity signals continued open-weights activity in the safety tooling space.
inclusionAI released Qwen3.5-0.8B-singprobe on Hugging Face under Apache-2.0. The probe reuses the base model's hidden states from layers [6, 14, 22] to output 8 intents plus unsafe and hallucination scores per token. Reported F1 is 0.8334 for query intent and response safety, R-AUC/T-AUC 0.9778/0.9206 for streaming safety, and AUC 0.7480 for hallucination detection. Benign-response false-positive rate is 0.23% average across 5 datasets. Training code is available at inclusionAI/SingProbe, and deployment is supported via SGLang and vLLM integration branches.
The probe adds 2.23M parameters and taps layers [6, 14, 22] of Qwen3.5-0.8B, producing per-token scores with less than 0.5% decode overhead. Reported metrics are company_reported; independent evaluation is not provided in the evidence.
Developers using Qwen3.5-0.8B can add streaming safety and hallucination scoring without a separate safety model, reducing deployment complexity and latency. The Apache-2.0 license and open training code lower the barrier for enterprises to adopt intrinsic guardrails.
For teams already running Qwen3.5-0.8B, this probe offers a low-overhead way to add safety and hallucination monitoring without additional model serving costs. The open license and code may reduce compliance and integration effort for enterprise deployments.
Next signals to check: independent benchmarks of the probe's F1/AUC against the cited baselines, adoption in SGLang/vLLM production deployments, and whether inclusionAI releases probes for larger Qwen3.5 models.