star-lab/STAR-0b6

Hugging Face
TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 21, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

STAR-0b6 is a 0.6 billion parameter causal language model developed by star-lab, fine-tuned from Qwen/Qwen3-0.6B. Specialized in function calling, it utilizes the STAR (Similarity-guided Teacher-Assisted Refinement) framework, which includes Similarity-guided RL and Constrained Knowledge Distillation. This model achieves strong performance on function calling benchmarks for its size, supporting a context length of up to 32,768 tokens.

Loading preview...

Overview

STAR-0b6 is a 0.6 billion parameter causal language model developed by star-lab, specifically fine-tuned for function calling. It is built upon the Qwen/Qwen3-0.6B base model and leverages the innovative STAR (Similarity-guided Teacher-Assisted Refinement) framework. This framework is designed to transfer advanced capabilities from larger LLMs into smaller models, making them efficient for agentic applications.

Key Innovations & Capabilities

The STAR framework introduces two core components:

  • Similarity-guided RL (Sim-RL): A reinforcement learning mechanism that uses a fine-grained, similarity-based reward signal for robust policy optimization, especially for complex, multi-solution tasks like function calling.
  • Constrained Knowledge Distillation (CKD): An advanced training objective that enhances stability and preserves exploration capacity by suppressing confidently incorrect predictions.

STAR-0b6 demonstrates strong performance on function calling benchmarks, significantly outperforming other open models under 1B parameters and even surpassing several larger models. It achieved 51.70% overall accuracy on BFCLv3 and a 53.00% summary score on ACEBench, indicating superior generalization and robustness compared to its base model's 27.20%.

Use Cases

  • Function Calling: Ideal for applications requiring models to accurately identify and execute external tools or APIs based on natural language prompts.
  • Agentic Applications: Suitable for building efficient AI agents that need to interact with their environment through function calls.
  • Resource-Constrained Environments: Its small parameter count (0.6B) combined with a substantial context length (32,768 tokens) makes it suitable for deployment in environments where computational resources are limited.