open-thoughts/OpenThinker-Agent-v1-SFT

OpenThinker-Agent-v1-SFT: Agentic Model for Complex Tasks

OpenThinker-Agent-v1-SFT, developed by OpenThoughts, is a supervised fine-tuned (SFT) model derived from the Qwen3-8B architecture. It is part of a broader effort to create high-performing open-source agents, with this specific release representing the SFT stage before further reinforcement learning (RL).

Key Capabilities

• Agentic Task Performance: Specifically trained for agentic tasks, demonstrating strong performance on benchmarks such as Terminal-Bench 2.0 and SWE-Bench.
• Dataset Foundation: Fine-tuned on the OpenThoughts-Agent-v1-SFT dataset, which includes approximately 15,200 traces from sources like nl2bash (shell command formatting) and InferredBugs (C# and Java bug fixing).
• Performance Improvement: Shows significant improvements over the base Qwen3-8B model on agent benchmarks, with a Terminal-Bench 2.0 score of 4.9 and SWE-Bench Verified score of 15.7.

Good For

• Developing Agentic Systems: Ideal for researchers and developers building or experimenting with AI agents that require robust problem-solving capabilities in environments like terminals or codebases.
• Foundation for Further Training: Serves as a strong base model for additional fine-tuning or reinforcement learning, particularly for agentic applications, before moving to the full OpenThinker-Agent-v1 RL model.