AXIOM-TECH/Ouro-2.6B-Thinking-IBNN

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:32kPublished:Aug 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AXIOM-TECH/Ouro-2.6B-Thinking-IBNN is a 2.6 billion parameter language model developed by AXIOM-TECH, designed for latent reasoning tasks. This model incorporates a custom architecture with configurable parameters like 'total_ut_steps' and 'early_exit_threshold' to enhance its thinking process. It is specifically optimized for tasks requiring iterative thought and complex problem-solving, leveraging a looped language model approach.

Loading preview...

Overview

AXIOM-TECH/Ouro-2.6B-Thinking-IBNN is a 2.6 billion parameter language model developed by AXIOM-TECH, focusing on advanced latent reasoning capabilities. This model is built upon a unique architecture that allows for configurable "thinking" steps, aiming to improve its ability to process and solve complex problems iteratively.

Key Features

  • Custom Architecture: Integrates specific parameters like total_ut_steps (set to 4) and early_exit_threshold (set to 1.0) to control its internal reasoning process.
  • Iterative Thinking: Designed to perform latent reasoning via a looped language model approach, as detailed in the associated research paper.
  • Optimized for Inference: Utilizes torch.compile with mode="max-autotune" and attn_implementation="sdpa" for efficient execution on CUDA-enabled devices.

Use Cases

This model is particularly suited for applications requiring:

  • Complex problem-solving.
  • Tasks benefiting from iterative thought processes.
  • Research into advanced reasoning mechanisms in LLMs.

Project Links

Finetunes
2 models