nexus-syntegra/Nexus-TinyFunction-1.2B-v2.0
Nexus-TinyFunction-1.2B-v2.0 by Nexus-Syntegra is a 1.2 billion parameter function-calling model, fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct. Utilizing a hybrid recurrent-attention architecture, it offers significantly faster inference than transformer-only models, enabling deployment on mobile and edge devices. This model excels at direct tool calls without verbose reasoning, achieving 94.25% accuracy in simple function calling and 99.3% JSON syntax reliability, making it ideal for efficient, low-latency function execution.
Loading preview...
Nexus-TinyFunction-1.2B-v2.0: Fast, Edge-Optimized Function Calling
Nexus-TinyFunction-1.2B-v2.0 is a 1.2 billion parameter model developed by Nexus-Syntegra, specifically fine-tuned for efficient function calling. Built upon LiquidAI's LFM2.5-1.2B-Instruct, it leverages a hybrid recurrent-attention architecture that enables blazing fast inference, outperforming many similarly sized (and even smaller) pure transformer models. This design allows it to run effectively on resource-constrained devices like mobile phones and Raspberry Pi, with a Q4_K_M quantization fitting in approximately 700 MB.
Key Capabilities
- Direct Tool Calls: Unlike reasoning-based models, it produces accurate tool calls directly, eliminating the need for chain-of-thought or verbose reasoning, which minimizes latency and token usage.
- High Accuracy: Achieves 94.25% accuracy in simple function calling and a remarkable 99.3% JSON syntax reliability, ensuring structured and valid outputs.
- Robust Irrelevance Detection: Demonstrates strong irrelevance detection (80.42%), reliably refusing to call tools when no match is found, preventing hallucinated function calls.
- Parallel & Multiple Tool Calling: Capable of handling complex scenarios involving multiple and parallel tool calls.
- Edge Deployment: Optimized for deployment on edge devices, including Android phones and Raspberry Pi, due to its efficient architecture and small footprint.
Use Cases
This model is particularly well-suited for applications requiring:
- Low-latency function execution in constrained environments.
- Reliable and accurate tool orchestration without excessive computational overhead.
- Edge AI applications where speed and minimal resource usage are critical.
It was trained using QLoRA SFT with a 3-stage curriculum learning approach, followed by Direct Preference Optimization (DPO), on a dataset of approximately 38,500 curated examples.