lausannequants/OpenThinkerAgent-8B-ColdStartSFTForRL

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The OpenThinkerAgent-8B-ColdStartSFTForRL is an 8 billion parameter Qwen3-based causal language model developed by OpenThoughts-Agent, featuring a 40,960-token context length. It serves as the cold-start, pre-Reinforcement Learning (RL) base model within the OpenThoughts-Agent SFT→RL pipeline. This model is specifically fine-tuned for agentic interaction formats and tool-use behavior, preparing it for subsequent RL training to create a fully capable agent.

Loading preview...

OpenThinkerAgent-8B-ColdStartSFTForRL Overview

This model, developed by OpenThoughts-Agent, is an 8 billion parameter Qwen3-based causal language model with a substantial 40,960-token context length. It represents a crucial step in the OpenThoughts-Agent SFT→RL (Supervised Fine-Tuning to Reinforcement Learning) pipeline, specifically designed as the cold-start, pre-RL base.

Key Characteristics and Purpose

  • Base Architecture: Built upon Qwen/Qwen3-8B, featuring a Qwen3ForCausalLM architecture with 36 layers and a hidden size of 4096.
  • Training Data: Fine-tuned using full-parameter SFT on the OpenThoughts-Agent-SFT-ColdStartForRL-10K dataset, which comprises nearly 9,500 sandboxed coding tasks with oracle-verified solutions.
  • Role in Pipeline: Its primary purpose is to instill the necessary agentic interaction format and tool-use conventions required for stable and effective subsequent reinforcement learning. It is intended to be further trained into the final OpenThinkerAgent-8B-RL model.

Intended Use and Limitations

This checkpoint is explicitly designed as a starting point for agentic RL and is not intended as a final, deployable agent on its own. While it has learned the foundational agentic behaviors, its standalone performance is expected to be lower than its RL-trained successor. Users should be aware that outputs may require review, and no standalone agentic benchmark numbers are provided for this cold-start version.