Amity Sigma Thinking v3r: Agentic Customer Service Model
Amity Sigma Thinking v3r is a 4-billion parameter language model developed by Amity, specifically engineered for high-performance agentic customer service. It is built upon the Qwen3-4B-Thinking base model and has undergone a specialized training regimen involving ROAD optimization and GRPO fine-tuning on synthetic customer service data with human-in-the-loop feedback.
Key Capabilities & Performance
- Exceptional Retail Customer Service: Achieves an impressive 78.51% Pass@1 on the Tau2-Bench retail domain, demonstrating strong capabilities in handling complex customer interactions.
- Outperforms Larger Models: Remarkably, it surpasses GPT-4.1 (1.7T parameters) by 4.5 percentage points and Qwen3-Max (340-407B parameters) by 3.1 percentage points on the Tau2-Bench retail domain, despite being significantly smaller.
- Cost-Efficient: Designed for self-hosting, offering estimated costs of ~$1-2 per 1,000 customer interactions, making it substantially more economical than commercial APIs like Claude Sonnet 4 or GPT-4.1.
- Tool-Calling: Optimized for tool-calling and function execution, crucial for agentic workflows.
Ideal Use Cases
- Customer Service Automation: Automating responses and actions in customer support.
- Order Management: Handling inquiries and tasks related to order processing.
- Multi-turn Conversational Agents: Building sophisticated chatbots capable of extended dialogues.
- Self-Hosting: Suitable for organizations prioritizing data privacy, predictable costs, and low-latency local inference on modest hardware (e.g., 1x A100 40GB or 2x RTX 4090).