trinityomni/OpenThinkerAgent-32B-SFT-10K

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OpenThinkerAgent-32B-SFT-10K is a 32 billion parameter language model developed by trinityomni, post-trained from Qwen3-32B. It is fine-tuned using 10,000 examples from the OpenThoughts-Agent-SFT-10K dataset, specifically optimized for agentic tasks. This model demonstrates enhanced performance on benchmarks like SWE-Bench-Verified-100, OpenThoughts-TBLite, and Terminal-Bench 2.0, making it suitable for complex problem-solving and automated agent workflows.

Loading preview...

OpenThinkerAgent-32B-SFT-10K Overview

OpenThinkerAgent-32B-SFT-10K is a 32 billion parameter model developed by trinityomni, built upon the Qwen3-32B architecture. This model is the result of a significant open-source effort to curate high-quality datasets for training agentic models, specifically leveraging the OpenThoughts-Agent-SFT-10K dataset.

Key Capabilities & Training

  • Agentic Task Optimization: The model is post-trained with full-parameter Supervised Fine-Tuning (SFT) on 10,000 examples from the OpenThoughts-Agent-SFT-10K dataset. This dataset comprises (task, agent-trajectory) pairs derived from sources like SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos, and IssueTasks.
  • Performance Improvement: It shows notable improvements over its base model, Qwen3-32B, across several agent-specific benchmarks. For instance, it achieves 35.0 on SWE-Bench-Verified-100 (compared to 26.7 for Qwen3-32B), 33.2 on OpenThoughts-TBLite (compared to 13.7), and 16.9 on Terminal-Bench 2.0 (compared to 7.5).
  • Data Generation: The training trajectories were generated by GLM-4.7-AWQ within the terminus-2 harness and filtered for traces with at least 5 model turns, ensuring high-quality, multi-turn interaction data.

When to Use This Model

  • Automated Agent Development: Ideal for applications requiring robust agentic capabilities, such as automated problem-solving, code generation, and complex task execution.
  • Research in Agentic AI: Useful for researchers exploring advanced agent architectures and data-driven training methodologies.
  • Benchmarking Agent Performance: Provides a strong baseline and improved performance for evaluating agentic systems against established benchmarks.

Finetunes
1 models