ZeroDrift Inc. has introduced Anchor 3.0, a suite of three small language models designed to automate compliance checks for AI-generated communications. The models are accessible via the company's Enforcement API and are engineered to inspect outgoing messages in real time to ensure adherence to financial regulations and corporate policies.
The Anchor 3.0 lineup includes three distinct versions. The Mini model features 9 billion parameters with 4 billion active, optimized for high-volume traffic. The flagship model supports over 200 prebuilt rules, including those from FINRA and the SEC, and can rewrite non-compliant text. The Max version, with 27 billion parameters, is designed for long documents and internal policy enforcement.
The models were post-trained using Google's Gemma E4B and Alibaba's Qwen3.8-27B architectures. According to internal benchmarks conducted by ZeroDrift using data from Surge AI Inc., the flagship model matched the accuracy of GPT-6 Astra and Claude Fable 5.1, while operating over 100 times faster and at less than one 500th of the cost.
The system integrates into the broader ZeroDrift platform, which can flag, block, or route messages for human review. The company also offers Guard for Agents, an API service that embeds these compliance checks directly into existing agent workflows. These tools aim to address the operational challenge of monitoring autonomous agents at scale.
