Qwen3-1.7B-FC: Function Calling Specialist

Qwen3-1.7B-FC is a 1.7 billion parameter model, fine-tuned from the Qwen3-1.7B base using Reinforcement Learning with Verifiable Rewards (RLVR). This specialized training significantly improves its function calling capabilities, making it highly efficient and accurate for tool-use scenarios. It achieves a 54.2% overall score on the BFCL V3 benchmark, outperforming the 8B and 14B parameter Qwen3 models, and generates 36% fewer tokens per response by producing direct tool calls and concise refusals.

Key Capabilities

• High Accuracy Function Calling: Surpasses larger models in core, executable Python, and live API function calling tasks.
• Efficient Responses: Optimized to generate direct tool calls without verbose preambles, reducing token count.
• Robust Refusal: Trained on 46,000 negative samples to effectively identify and refuse queries that cannot be answered by available tools, minimizing hallucinations.
• Compact Size: At 1.7B parameters, it is suitable for deployment on consumer-grade GPUs and edge devices.
• Reduced Catastrophic Forgetting: RLVR training helps maintain general chat abilities and multilingual support (English and Vietnamese).

Ideal Use Cases

• Edge Device Deployment: Runs efficiently on devices with limited resources.
• Customer Service Automation: Automates tasks like order lookup, ticket creation, and FAQ responses via tool calls.
• Voice Agents / Call Centers: Enables real-time voice-to-action systems.
• IoT/Smart Home: Controls devices through function calling on local hardware.
• Cost-Efficient API Gateway: Routes requests to backend services with low latency and cost.