Skip to content

Light-R1

An open project for training long chain-of-thought models

Visit official site

What Light-R1 is

The project is an open-source effort from Qihoo360 for training long chain reasoning models. It uses curriculum SFT, DPO and RL with decontaminated data and open code.

What you can do with it

  • Train long chain reasoning from scratch
  • Use the open dataset and training code
  • Run SFT scripts and then DPO scripts

Getting started

  • Clone the Light-R1 code locally
  • Install the required Python packages
  • Run SFT training with the open dataset, then DPO

Pros & cons

✓ What we like

  • AIME24 and AIME25 benchmarks are named
  • The material describes low-cost training

! What to watch out for

  • The named focus is math reasoning
  • Training needs a Python environment and dependencies

FAQ

What accuracy figures are listed?

The material lists 59.1 percent for Light-R1-7B-DS and 76.6 percent for Light-R1-32B on AIME24.

Who developed it?

The material names Qihoo360.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review