inclusionAI releases SingProbe streaming safety probes for Qwen and GLM Flash models

1. 12:47 AM
   inclusionAI (Ant Ling) · new models on Hugging FaceOfficial

   inclusionAI releases GLM-5.3-Flash-singprobe, a streaming safety probe for GLM-5.3-Flash

   AIinclusionAI releases GLM-5.3-Flash-singprobe, an 8.13M-parameter probe built on zai-org/GLM-5.3-Flash that scores query intent, response unsafety, and hallucination risk at every generated token. The probe reuses the base model's hidden states from layers 14, 29, and 43 and adds less than 0.5% decode-time overhead. The model card reports a 0.02% benign-response false-positive rate averaged across five datasets, and the probe loads through SGLang or vLLM integration branches.

2. 12:46 AM
   inclusionAI (Ant Ling) · new models on Hugging FaceOfficial

   inclusionAI releases SingProbe, a lightweight streaming guardrail for Qwen3.8-Flash-Next

   AIinclusionAI releases SingProbe, a 5.18M-parameter probe on Qwen/Qwen3.8-Flash-Next that scores query intent, response unsafety, and hallucination risk at every generated token using the base model's hidden states. The probe adds less than 0.5% decode-time overhead and posts an 0.8714 F1 on response safety classification across eight benchmarks, against 0.8604 for Qwen3Guard-Gen-8B-strict. Training code is available at inclusionAI/SingProbe, and the probe loads by Hugging Face ID in SGLang or vLLM.

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.