DeepScaleR-1.5B-Preview
Open source math modelOpen source math model for AIME and AMC
Highlights
What you can do with it
- Plug it into math competition tutoring software for step by step hints.
- Study the open source code while exploring reinforcement learning methods.
- Generate practice problems and solutions for classroom use.
How to use it
- 1 Download the model files from Hugging Face.
- 2 Install a supported inference system such as vLLM or Hugging Face Text Generation Inference.
- 3 Load the model and set parameters such as context length and sampling strategy.
- 4 Call the model through its API to solve math problems.
- 5 Parse the output to extract the answer or the solution steps.
Who it is for
Pros
- MIT licensed and free for researchers and developers.
- Strong focus on math, including competition level questions.
- Open architecture and training methods make it easy to build on.
- A long 24K token context suits multi step problems.
Limitations
- It is a preview at 1.5B parameters, so treat it as a research artifact rather than a finished product.
- Setting it up needs an inference system plus GPU and engineering work.
- Math answers can still be wrong and need checking.
- Open source licensing reflects the current release and can change in later versions.
What DeepScaleR-1.5B-Preview is
DeepScaleR-1.5B-Preview is a large language model tuned with reinforcement learning to improve mathematical problem solving. It comes from the Sky Computing Lab and the Berkeley AI Research team at the University of California, Berkeley, and it is released under the MIT license, which means researchers and developers can use it at no cost. The team reports accuracy gains in long text reasoning, driven by distributed reinforcement learning algorithms. It reads contexts up to 24K tokens and was trained on a large dataset of math problems that includes AIME and AMC competition questions. The stated aim is to push the use of AI in education, especially in math teaching and competitive math.
What you can do with it
- Run it locally to solve math problems through an API.
- Build it into tutoring software that gives students hints and solution steps.
- Read the open architecture and training code to study reinforcement learning methods.
- Generate practice problems and worked solutions for a class.
Who it is for
- Math teachers and tutors who need fresh practice material.
- Researchers exploring reinforcement learning for reasoning.
- Developers building math and problem solving tools.
What to watch out for
This is a preview release of a 1.5 billion parameter model, so treat it as a research artifact rather than a polished product. Running it means downloading the weights, installing a supported inference system such as vLLM or Hugging Face Text Generation Inference, and setting parameters yourself, which needs engineering time and hardware. Even with its math focus, the model can produce wrong answers, so any result used for teaching should be checked by a person. Licensing and availability reflect the current release and can change in later versions, and the reference page was not reachable during the last automated check.
FAQ
Is the model free?
Yes. It is released under the MIT license and is free for researchers and developers.
Who built it?
The Sky Computing Lab and the Berkeley AI Research team at the University of California, Berkeley.
What is it good at?
Mathematics. It was trained on a large math dataset that includes AIME and AMC competition questions.
What do I need to run it?
A supported inference system such as vLLM or Hugging Face Text Generation Inference, plus the model files from Hugging Face.
Why it is free
What the official page says： 完全免费免费
Free limits, trial length and features are decided by the official site and can change at any time.
Product preview
Screenshot archived from the listing page, for preview only.
