inclusionAI releases a streaming guardrail for DeepSeek-V4.1-Flash-1017
Model: DeepSeek-V4.1-Flash-1017, DeepSeek-V4.1-Flash · availability, license and releases
inclusionAI released DeepSeek-V4.1-Flash-1017-singprobe, a 10.10M-parameter guardrail that scores query intent, response unsafety, and hallucination risk at each token. It reads hidden states from three layers of the base model. The bfloat16 weights are available on Hugging Face under the MIT license.
China context
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the probe weights from Hugging Face and test them with the matching DeepSeek base model. The model card does not establish whether a hosted API exposes the probe's per-token scores.
SingProbe uses an attention head to read DeepSeek-V4.1-Flash-1017's hidden states during generation and score eight query intents, response unsafety, and hallucination risk. The model card documents an SGLang setup that exposes these scores in generation responses.
inclusionAI reports streaming safety AUCs of 0.9907 at the response level and 0.9522 at the token level, both above the listed baseline. Its hallucination detection AUC was 0.7951, below the DRIFT baseline of 0.8000. The reported 0.009% false-positive rate used a pre-calibration threshold and does not establish the rate at the release threshold of 0.5.
Teams serving DeepSeek-V4.1-Flash-1017 can use the released probe to score response unsafety and hallucination risk before generation finishes. The documented setup uses a specific SGLang build and eight H20 GPUs. The release does not establish the probe's incremental serving cost or how often score-based interventions would interrupt benign responses.
The MIT-licensed weights give teams already running the supported base model a way to test token-level monitoring in their own stack. Before deploying it, teams can measure serving overhead and compare error rates with those of their existing moderation system.
The next checks are independent measurements of latency and throughput with the probe enabled, and false-positive and false-negative rates at the shipped thresholds on workloads outside the release benchmarks.