open-thoughts/OpenThinker-Agent-v1

OpenThinker-Agent-v1 by open-thoughts is an 8 billion parameter language model, post-trained from Qwen3-8B, specifically optimized for agentic tasks. It excels in environments like Terminal-Bench 2.0 and SWE-Bench, demonstrating state-of-the-art performance at its scale on these agent benchmarks. The model was developed using a two-stage process involving supervised fine-tuning (SFT) and reinforcement learning (RL) on curated datasets.