ASTRA-32B-Thinking-v1 Overview

ASTRA-32B-Thinking-v1 is a 32 billion parameter model, based on Qwen3-32B, designed for advanced agentic capabilities. It specializes in multi-step, tool-augmented tasks, demonstrating enhanced performance in complex tool use and structured reasoning. The model's development is detailed in the paper "ASTRA: Automated Synthesis of agentic Trajectories and Reinforcement Arenas".

Key Capabilities

• Optimized for Agentic Tasks: Specifically fine-tuned for scenarios requiring sophisticated multi-step reasoning and tool integration.
• Enhanced Tool Use: Leverages an extensive tool pool (19,036 tools across 41 domains) for generating realistic and parameter-satisfiable tool-chains.
• Reinforcement Learning with Verifiable Rewards: Utilizes automated verifiable environments synthesized in Python, providing multi-turn, step-wise verifiable training signals.
• Strong Performance: Achieves state-of-the-art results on the BFCL-V3 multi-turn subset at comparable model scales.

Good for

• Developing AI agents that require complex, multi-step decision-making.
• Applications involving tool-augmented reasoning and automated task execution.
• Research into agentic AI, reinforcement learning with verifiable environments, and automated trajectory synthesis.