inclusionAI released GLM-5.3-Flash-singprobe, a streaming safety probe for GLM-5.3-Flash
inclusionAI released GLM-5.3-Flash-singprobe, an intrinsic streaming guardrail built on zai-org/GLM-5.3-Flash, on Hugging Face. It adds less than 0.5% decode-time overhead and has a 0.02% benign-response false-positive rate.
China context
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can integrate the probe via SGLang or vLLM branches to add streaming safety scoring to GLM-5.3-Flash deployments.
- For investors
- Investors can monitor adoption of the probe as a signal for demand in lightweight safety tooling for open-weight models.
inclusionAI released GLM-5.3-Flash-singprobe on Hugging Face. The model is an intrinsic streaming guardrail built on zai-org/GLM-5.3-Flash 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 has a 0.02% benign-response false-positive rate. Training codes are available at inclusionAI/SingProbe.
The probe has 8.13M parameters, taps layers [14, 29, 43], and outputs 8 intents plus unsafe and hallucination scores. Reported F1 scores are 0.8683 for query intent classification, 0.8706 for response safety classification, and AUC 0.8143 for hallucination detection. Streaming safety R-AUC/T-AUC are 0.9786/0.9232.
Developers using GLM-5.3-Flash can add streaming safety scoring with minimal overhead, avoiding a separate safety model. This reduces integration cost and latency for real-time guardrails.
The probe offers a low-overhead safety layer for GLM-5.3-Flash deployments, potentially reducing the need for separate safety models and lowering operational costs.
Adoption can be checked via downloads and community usage on Hugging Face, and whether the SGLang/vLLM integration branches are merged upstream.