Stage-org/4b-300-LH-27b-z-iter2-epoch3

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026Architecture:Transformer Featherless Exclusive Cold

The Stage-org/4b-300-LH-27b-z-iter2-epoch3 model is a 4.5 billion parameter language model developed by Stage-org, trained using a reinforcement learning (RL) method. It is specifically optimized for complex reasoning tasks within an agent-based environment, leveraging a large sequence length of 300,000 tokens. This model is designed for applications requiring advanced decision-making and problem-solving capabilities, particularly in scenarios involving open-ended judgment and strategic generation.

Loading preview...

Model Overview

Stage-org/4b-300-LH-27b-z-iter2-epoch3 is a 4.5 billion parameter language model developed by Stage-org, distinguished by its training methodology focused on reinforcement learning (RL). This model was trained on the Stage-org/4b-300-LH-27b-z-iter2 dataset, emphasizing iterative learning and refinement through an agent-based approach.

Key Capabilities

  • Reinforcement Learning (RL) Optimization: The model's core strength lies in its RL-driven training, enabling it to learn and adapt within dynamic environments.
  • Extended Context Handling: It supports a substantial sequence length of 300,000 tokens, facilitating the processing of extensive inputs and complex scenarios.
  • Advanced Reasoning and Judgment: Incorporates an "open-ended judge" mechanism, utilizing models like gpt-5.6-luna for sophisticated evaluation and decision-making, with features like reasoning effort and retry logic.
  • Strategic Generation: Designed for generation tasks with configurable parameters such as temperature and top_p, and includes an "enable_thinking" feature for more deliberate output.
  • Specialized Parsing: Utilizes qwen3 for reasoning parsing and qwen3_coder for tool call parsing, indicating potential for structured output and code-related tasks.

Good For

  • Agent-based Systems: Ideal for integration into AI agents requiring complex decision-making and interaction within simulated or real-world environments.
  • Problem Solving: Suitable for tasks that benefit from iterative refinement and strategic planning, leveraging its RL foundation.
  • Long-Context Applications: Excels in scenarios where understanding and generating responses based on very long input sequences are critical.
  • Research in RL and LLMs: Provides a robust platform for exploring advanced reinforcement learning techniques applied to large language models, particularly concerning open-ended judgment and strategic generation.