Robots Atlas>ROBOTS ATLAS
Qwen-2.5-7B-SimpleRL-Zoo
AI Models

Qwen-2.5-7B-SimpleRL-Zoo

Qwen2.5-7B · SimpleRL-Zoo
A 7B reasoning model: base Qwen2.5-7B trained with 'zero RL' (RLVR/GRPO) without SFT, from HKUST NLP's SimpleRL-Zoo. Boosted mathematical reasoning.
🔬 Research✓ Public access⚖ Open sourceLLMReasoning model
Parameters
7B
parameters
Release date
24 March 2025
🏢HKUST NLPProducer
Access:DownloadDeployment:💻 Local

Overview

Qwen-2.5-7B-SimpleRL-Zoo is a reasoning model produced by applying 'zero reinforcement learning' directly to the base Qwen2.5-7B model, without prior instruction tuning (SFT). It comes from the SimpleRL-Zoo project by the HKUST NLP group (led by Junxian He).

SimpleRL-Zoo investigates and tames zero-RL for open base models 'in the wild', using strategies such as format-reward adjustment and query-difficulty control to strengthen reasoning. The study found that different base models train differently and, for the first time, observed the emergence of verification behaviors (the 'aha moment') in small non-Qwen models.

Training used reinforcement learning with verifiable rewards (RLVR) and GRPO optimization. The model was evaluated on mathematical reasoning benchmarks (AIME, MATH500, GSM8K). Weights are released under Apache 2.0. The work was presented at COLM 2025 (arXiv:2503.18892).

Classification
LLMReasoning model
Access & deployment
Download
Local
Weights: Open source
Key parameters
🧩 Parameters: 7B
✓ Fine-tuning
📥 Input: text

Technical specification

Parameters
7B
parameters
License
Apache 2.0
Hardware requirements
Base Qwen2.5-7B after zero-RL training (RLVR/GRPO), without SFT.
Features:✓ Fine-tuning
Modalities
⬇ Input
text
⬆ Output
textcode

Capabilities and applications

Native model capabilities
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning — equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language

Benchmark results

3 benchmarks
AIME
mathematical reasoning
📄 paper
Evaluated per the SimpleRL-Zoo paper (COLM 2025).
MATH500
📄 paper
Evaluated per the SimpleRL-Zoo paper.
GSM8K
📄 paper
Evaluated per the SimpleRL-Zoo paper.

Technical architecture

Core Architecture
TRTransformer
Training Techniques
RLRLVRGRGRPORLRL

Sources and related pages