inclusionAI open-sourced MiniMax-M2.7-singprobe, a streaming guardrail probe for MiniMax-M2.7
Model: MiniMax-M2.7, MiniMax-M2 · availability, license and releases
inclusionAI released MiniMax-M2.7-singprobe, a streaming guardrail probe built on MiniMaxAI/MiniMax-M2.7, 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
- 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 to add streaming safety scoring to MiniMax-M2.7 deployments.
- For investors
- The release signals inclusionAI's focus on integrated safety tooling for its models, which may differentiate its offerings in markets requiring compliance.
inclusionAI released MiniMax-M2.7-singprobe on Hugging Face. The probe reuses the base model's hidden states during generation to score query intent, response unsafety, and hallucination risk at every token, adding less than 0.5% decode-time overhead. It is supported through SGLang and vLLM integration branches.
The probe taps layers [19, 39, 60] of MiniMax-M2.7 and outputs 8 intents plus unsafe and hallucination scores. Reported F1 is 0.8731 for query intent classification, 0.8698 for response safety, and AUC 0.7777 for hallucination detection. Benign-response false-positive rate is 0.03% across 5 datasets.
Developers using MiniMax-M2.7 can add streaming safety and hallucination scoring without deploying a separate guard model, reducing infrastructure cost and latency. Competitors offering external guard models face pressure to match this integrated approach.
The probe lowers the cost and complexity of adding safety and hallucination detection to MiniMax-M2.7 deployments, potentially increasing enterprise adoption of the base model.
Next signals to check: adoption of the SGLang/vLLM integration branches, independent evaluation of the reported metrics, and whether inclusionAI extends the probe to other base models.