open-thoughts/OpenThinkerAgent-32B-SFT-100K

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OpenThinkerAgent-32B-SFT-100K is a 32 billion parameter language model developed by OpenThoughts, fine-tuned from Qwen3-32B. It is specifically optimized for agentic tasks, leveraging a 100,000-example supervised fine-tuning dataset of agent trajectories. This model excels in complex problem-solving environments, demonstrating significant performance improvements on benchmarks like SWE-Bench and Terminal-Bench.

Loading preview...

OpenThinkerAgent-32B-SFT-100K Overview

OpenThinkerAgent-32B-SFT-100K is a 32 billion parameter model developed by OpenThoughts, built upon the Qwen3-32B architecture. This model is the result of an open-source initiative focused on curating high-quality datasets for training AI agents.

Key Capabilities and Training

This model has undergone full-parameter supervised fine-tuning (SFT) using the extensive OpenThoughts-Agent-SFT-100K dataset. This dataset comprises 100,000 examples of (task, agent-trajectory) pairs, sourced from platforms like SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos, and IssueTasks. The trajectories were generated by a GLM-4.7-AWQ teacher model and filtered for traces with at least five turns, ensuring high-quality, multi-step reasoning examples.

Performance Highlights

Evaluated within the terminus-2 harness, OpenThinkerAgent-32B-SFT-100K demonstrates substantial performance gains over its base model, Qwen3-32B, particularly in agentic benchmarks:

  • SWE-Bench-Verified-100: Achieves 55.7%, significantly up from 26.7%.
  • OpenThoughts-TBLite: Scores 41.3%, compared to 13.7%.
  • Terminal-Bench 2.0: Reaches 26.2%, an improvement from 7.5%.

Ideal Use Cases

This model is particularly well-suited for applications requiring advanced agentic capabilities, such as:

  • Automated code generation and debugging: Excels in environments like SWE-Bench.
  • Complex problem-solving: Designed for tasks requiring multi-step reasoning and interaction within terminal-like environments.
  • Agent development: Provides a strong foundation for building and deploying AI agents that can autonomously complete intricate tasks.

Finetunes
1 models