inclusionAI open-sourced SingProbe, a streaming safety probe for Qwen3.5-122B-A10B
inclusionAI released Qwen3.5-122B-A10B-singprobe on Hugging Face under apache-2.0. It 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.
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 via SGLang or vLLM branches to add streaming safety scoring to Qwen3.5-122B-A10B deployments.
- For investors
- The release of an open-weights safety probe by an Ant Group-affiliated entity signals continued investment in model safety tooling, which may influence enterprise adoption of Chinese open models.
inclusionAI released Qwen3.5-122B-A10B-singprobe on Hugging Face under apache-2.0. The probe is built on Qwen/Qwen3.5-122B-A10B and adds 6.17M parameters tapped at layers [14, 30, 46], outputting 8 intents plus unsafe and hallucination scores. Evaluation results show F1 0.8697 for query intent classification, F1 0.8680 for response safety classification, R-AUC/T-AUC 0.9895/0.9304 for streaming safety, and AUC 0.7806 for hallucination detection. Benign-response false-positive rate is 0.03% average across 5 datasets. Training codes are available at inclusionAI/SingProbe, and integrations are provided via SGLang and vLLM branches.
SingProbe is a lightweight probe (6.17M parameters) that taps hidden states from layers [14, 30, 46] of Qwen3.5-122B-A10B to produce per-token scores for 8 intents, unsafety, and hallucination. It adds less than 0.5% decode-time overhead. Reported metrics: query intent F1 0.8697, response safety F1 0.8680, streaming safety R-AUC/T-AUC 0.9895/0.9304, hallucination AUC 0.7806, benign false-positive rate 0.03%. These are company-reported; independent evaluation is not provided.
Developers using Qwen3.5-122B-A10B can add streaming safety and hallucination scoring without deploying a separate safety model, reducing inference cost and complexity. The probe's low overhead and Apache-2.0 license may accelerate adoption of intrinsic guardrails in production systems.
For enterprises running Qwen3.5-122B-A10B, SingProbe offers a low-latency, integrated safety layer that could lower operational costs compared to separate safety models. The open-source license allows commercial use without additional licensing fees.
Next observable signals include independent benchmarks of SingProbe against external guardrails, adoption in SGLang/vLLM deployments, and whether inclusionAI releases probes for other base models.