FRPO/qwen3-1.7b-a8_klreward-krew-k1-coef1e-4-mb4-eta100-bs256x5-n2

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026Architecture:Transformer Featherless Exclusive Cold

FRPO/qwen3-1.7b-a8_klreward-krew-k1-coef1e-4-mb4-eta100-bs256x5-n2 is a 2 billion parameter language model based on the Qwen3-1.7B architecture, developed by FRPO. This model is a checkpoint from KL-in-LLM-RL / FRPO experiments, specifically fine-tuned using Reinforcement Learning (RL) with the verl framework. It is designed for tasks benefiting from RL-based optimization, offering a 32768 token context length.

Loading preview...

Model Overview

FRPO/qwen3-1.7b-a8_klreward-krew-k1-coef1e-4-mb4-eta100-bs256x5-n2 is a 2 billion parameter language model derived from the Qwen3-1.7B base architecture. It represents a specific checkpoint from the KL-in-LLM-RL / FRPO experimental series, where it underwent Reinforcement Learning (RL) fine-tuning using the verl framework.

Key Characteristics

  • Base Model: Qwen/Qwen3-1.7B, indicating its foundational architecture.
  • Fine-tuning Method: Utilizes Reinforcement Learning (RL) with KL-reward, suggesting optimization for specific reward signals.
  • Checkpoint: The repository contains the global_step_200 checkpoint, representing a specific stage of the RL training process.
  • Weights: Provided in fp32 safetensors format, directly as saved by the trainer without further post-processing.
  • Context Length: Supports a substantial context window of 32768 tokens.

Intended Use Cases

This model is particularly suited for research and development in:

  • RL-based Language Model Applications: Exploring the effects and benefits of KL-reward-based RL fine-tuning.
  • Experimental Setups: Serving as a foundation for further experimentation with RL techniques on Qwen3-1.7B.
  • Performance Analysis: Evaluating the impact of specific RL configurations (encoded in the repo name) on model behavior and output quality.