sagnik3788/Qwen2.5-1.5B-Instruct-FC

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

sagnik3788/Qwen2.5-1.5B-Instruct-FC is a 1.5 billion parameter Qwen2.5 model, fine-tuned by sagnik3788, specifically for reliable function call generation. It achieves 85.65% on the BFCL v4 Non-Live AST benchmark, outperforming many larger models, including 72B and 685B parameter giants. This model is optimized for single-turn function calling, making it highly practical for edge devices and low-cost agent applications requiring precise tool use.

Loading preview...

Model Overview

sagnik3788/Qwen2.5-1.5B-Instruct-FC is a specialized 1.5 billion parameter model based on the Qwen2.5 architecture, meticulously fine-tuned for high-fidelity function call generation. Its primary distinction lies in its exceptional performance on the Berkeley Function-Calling Leaderboard (BFCL) v4 Non-Live AST benchmark, where it scores 85.65%. This places it ahead of 71 out of 109 models, including significantly larger models like DeepSeek-V3.2-Exp (685B) and Qwen2.5-72B, and even GPT-4o.

Key Capabilities

  • Superior Function Calling: Achieves 85.65% accuracy on single-turn non-live function calling, making it a top performer in its class and competitive with much larger models.
  • Compact Size: At 1.5 billion parameters, it's highly efficient for deployment on edge devices or in cost-sensitive applications.
  • Reliable Tool Use: Specifically trained to output {"name":..., "arguments":...} within <tool_call></tool_call> tags with high accuracy.
  • Qwen2.5 Base: Leverages the robust Qwen2.5-1.5B-Instruct as its foundation.
  • Merged Weights: Provided as merged bf16 weights, eliminating the need for PEFT during inference.

Training Details

The model was fine-tuned using QLoRA SFT on 10,000 verified synthetic function-calling examples generated via a 3-stage APIGen pipeline. This focused training ensures its specialization in single-turn function calling. The training utilized Unsloth 4-bit quantization for efficiency, completing 3 epochs.

Limitations

  • Single-Turn Only: Not designed for multi-turn conversations, memory, or web search tasks.
  • Synthetic Data: Performance may vary with real-world, domain-specific API schemas not represented in the training data.
  • No Irrelevance Training: May attempt a tool call even when no tool is appropriate; requires an external no_tool threshold for robust agent design.

Ideal Use Cases

  • Edge AI Agents: Its small size and high accuracy make it suitable for function calling on resource-constrained devices.
  • Cost-Effective Tool Use: Provides high-performance function calling at a significantly lower computational cost than larger models.
  • Specialized Agentic Workflows: Excellent for applications requiring precise, single-turn tool invocation, such as data extraction, simple task automation, or API interaction.

Finetunes
1 models