Skip to content

Huginn-0125

A recurrent depth model with 3.5B parameters

Visit official site

What Huginn-0125 is

The model is a latent recurrent depth model from the Tom Goldstein lab with 3.5 billion parameters trained on 800 billion tokens. It adjusts its compute at test time through recurrent depth.

What you can do with it

  • Generate code and logic reasoning results
  • Explore recurrent depth efficiency as a research tool

Getting started

  • Download the model with the transformers library
  • Set num_steps to adjust depth and use bfloat16 precision

Pros & cons

✓ What we like

  • bfloat16 mixed precision inference is supported
  • Guides and code examples are provided for developers

! What to watch out for

  • Users load the model through the transformers library
  • Advanced features are enabled through specific parameters

FAQ

How large is the model?

The card states 3.5 billion parameters and 800 billion training tokens.

How is inference run?

The card describes bfloat16 precision with a generate call and a num_steps setting.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review