konizquants/OpenThinker-Agent-v1-SFT

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The OpenThinker-Agent-v1-SFT model by OpenThoughts is an 8 billion parameter language model, post-trained from Qwen3-8B. It is specifically fine-tuned for agentic tasks, excelling in environments like Terminal-Bench 2.0 and SWE-Bench. This model represents the supervised fine-tuning (SFT) stage of the OpenThinker-Agent-v1 development, focusing on curating high-quality datasets for agent training.

Loading preview...

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

OpenThinker-Agent-v1-SFT is an 8 billion parameter model developed by OpenThoughts, derived from the Qwen3-8B architecture. This model is the result of the supervised fine-tuning (SFT) stage within the broader OpenThinker-Agent-v1 project, which aims to create robust models for agentic tasks.

Key Capabilities & Training

  • Agentic Task Specialization: Fine-tuned for performance on complex agentic benchmarks such as Terminal-Bench 2.0 and SWE-Bench.
  • Supervised Fine-Tuning (SFT): Trained on the OpenThoughts-Agent-v1-SFT dataset, which comprises approximately 15,200 traces. This dataset includes tasks like nl2bash (shell command formatting) and InferredBugs (C# and Java bug-fixing tasks).
  • Data Curation: The OpenThoughts project emphasizes curating high-quality datasets for agent training, utilizing a three-stage filtration pipeline to ensure data stability and quality.

Performance Context

While this specific model is the SFT stage, the full OpenThinker-Agent-v1 (which includes an additional Reinforcement Learning stage) demonstrates significant improvements over its base model, Qwen3-8B, on agent benchmarks. For instance, the full RL-trained model achieves 15.7% on SWE-Bench Verified compared to 0.7% for Qwen3-8B.

Use Cases

This model is particularly well-suited for developers and researchers focused on:

  • Developing AI agents: As a foundational SFT model for agentic workflows.
  • Automated code generation and debugging: Especially for tasks involving shell commands and bug resolution in programming languages.
  • Research in agentic AI: Providing a strong base for further experimentation and reinforcement learning.

Finetunes
1 models