zurichquants/OpenThinkerAgent-8B-ColdStartSFTForRL

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

OpenThinkerAgent-8B-ColdStartSFTForRL is an 8 billion parameter Qwen3-based causal language model developed by OpenThoughts. It serves as the cold-start, pre-reinforcement learning (RL) base for agentic models, fine-tuned with full-parameter SFT on the OpenThoughts-Agent-SFT-ColdStartForRL-10K dataset. This model is designed to establish agentic interaction formats and tool-use behaviors, preparing it for subsequent RL training to create a fully capable agent. It features a 40,960-token context length and is optimized for developing agentic AI systems.

Loading preview...

OpenThinkerAgent-8B-ColdStartSFTForRL: A Foundation for Agentic AI

This model, developed by OpenThoughts, is a Qwen3-8B architecture with 8 billion parameters and a substantial 40,960-token context length. It represents a crucial cold-start, pre-RL base in the OpenThoughts-Agent SFT→RL recipe for building agentic models.

Key Capabilities & Purpose

  • Agentic Interaction Format: It has been fine-tuned using full-parameter Supervised Fine-Tuning (SFT) on the OpenThoughts-Agent-SFT-ColdStartForRL-10K dataset.
  • Tool-Use Behavior: The SFT process specifically imbues the model with the necessary conventions for tool-use, essential for agentic operations.
  • RL Preparation: Its primary role is to provide a stable foundation for subsequent reinforcement learning, leading to the final OpenThinkerAgent-8B-RL model.

Training Details

  • Base Model: Built upon Qwen/Qwen3-8B.
  • Dataset: Trained on 9,437 (task, trajectory) pairs from the OpenThoughts-Agent-SFT-ColdStartForRL-10K dataset, which includes sandboxed coding tasks solved by a teacher model.
  • Procedure: Full-parameter SFT using LLaMA-Factory, with a learning rate of 4e-05 over 7 epochs.

Intended Use & Limitations

This checkpoint is explicitly designed as a starting point for agentic RL and is not intended as a standalone, deployable agent. Its agentic performance is expected to be lower than its RL-trained successor. Users should be aware that outputs may be incorrect or unsafe and require review, consistent with the base Qwen3-8B model.