Open-Reasoner-Zero/Open-Reasoner-Zero

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

2099 stars · 120 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Open-Reasoner-Zero is an open-source implementation of large-scale reinforcement learning training for reasoning-oriented language models, released with code, hyperparameters, training data, and model weights. It reproduces and improves upon DeepSeek-R1-Zero-style RL training, achieving strong benchmark results with fewer training steps.

Use cases

• train a reasoning LLM with reinforcement learning from a base model
• reproduce DeepSeek-R1-Zero style RL training pipeline
• scale up RL training across model sizes from 0.5B to 32B
• improve math reasoning benchmarks like AIME2024 and MATH500
• research minimalist RL recipes for LLM reasoning
• download pretrained reasoning model weights for fine-tuning

When to choose

• you want an open, reproducible RL training pipeline for reasoning models
• you have GPU cluster resources to train or fine-tune large language models
• you need training data, hyperparameters, and weights released together for research
• you want to study scalability of RL on base models

When to avoid

• you only need inference of a reasoning model without training
• you lack multi-GPU infrastructure for large-scale RL training
• you need a general-purpose RL library unrelated to LLMs
• you want a plug-and-play chat model with no training involved

Facets

library · maturity active

reinforcement-learning llm-training machine-learning reinforcement-learning large-language-models deep-learning machine-learning python reasoning rl-training llm open-source-research deepseek-r1 model-training gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Open-Reasoner-Zero/Open-Reasoner-Zero")