KhanCold/qwen3-8b-spader

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KhanCold/qwen3-8b-spader is an 8 billion parameter Qwen3-based model developed by Qiming Shi et al. It integrates the SPADER (Step-wise Peer Advantage with Diversity-Aware Exploration Rewards) reinforcement learning framework, specifically designed for long-horizon tool use in Multi-Answer Question Answering. This model excels at complex QA tasks by employing a critic-free credit assignment mechanism and a diversity-aware exploration reward to discover rare entities and avoid redundancy.

Loading preview...

Overview

KhanCold/qwen3-8b-spader is an 8 billion parameter model built upon the Qwen3 architecture, developed by Qiming Shi and colleagues. Its core innovation lies in the integration of the SPADER (Step-wise Peer Advantage with Diversity-Aware Exploration Rewards) reinforcement learning framework. This framework is specifically tailored for long-horizon tool use in Multi-Answer Question Answering (QA), addressing the challenges of complex information retrieval and synthesis.

Key Capabilities

  • Step-wise Peer Advantage (SPA): Utilizes a critic-free, step-level credit assignment mechanism. This aligns parallel trajectories by decision step and estimates advantages from peer returns, enhancing the model's ability to learn from multi-step processes.
  • Diversity-Aware Exploration Reward: Promotes the discovery of long-tail entities by dynamically upweighting rare findings and downweighting redundant ones. This encourages comprehensive and non-repetitive answer generation in multi-answer scenarios.
  • Multi-Answer Question Answering: Optimized for tasks requiring the identification and synthesis of multiple correct answers, often involving complex tool interactions.

Good For

  • Complex QA Systems: Ideal for applications that demand detailed, multi-faceted answers from diverse information sources.
  • Tool-Use Agents: Suitable for developing agents that need to perform a sequence of actions or use external tools to arrive at a comprehensive solution.
  • Research in Reinforcement Learning for LLMs: Provides a strong baseline and framework for further exploration into credit assignment and exploration strategies in language models.