Skip to content

DeepScaleR-1.5B-Preview

A maths-focused model tuned with reinforcement learning

Visit official site

What DeepScaleR-1.5B-Preview is

The model is described as a large language model optimised with reinforcement learning to improve mathematical problem solving, developed by Sky Computing Lab and Berkeley AI Research.

What you can do with it

  • Apply distributed reinforcement learning optimisation
  • Work with a long context of up to 24K
  • Draw on maths datasets covering AIME and AMC problems
  • Use an open architecture and training method for further work

Application scenarios

  • Add the model to contest tutoring software for live solving hints
  • Explore new reinforcement learning methods with the open code
  • Generate practice problems and solutions for classroom teaching

Getting started

  • Download the model files from Hugging Face
  • Install a supported inference system such as vLLM
  • Load the model and set context length and sampling options
  • Call the model service through an API for maths reasoning

Pros & cons

✓ What we like

  • Efficient inference support compatible with several high-performance systems
  • The MIT licence makes the model free for researchers and developers

! What to watch out for

  • The described focus is mathematics problem solving
  • Usage requires loading the model into a supported inference system

FAQ

Who developed it?

The material names Sky Computing Lab and Berkeley AI Research at UC Berkeley.

What licence applies?

The material states the model uses the MIT open-source licence.

Last reviewed: 2026-11-15

More AI education tools

View all →

How we review