omar81939/Ouro-1.4B-Thinking-depth-GRPO

TEXT GENERATIONPricing:Input $0.04 / Cached $0.002 / Output $0.08Concurrent Unit Cost:1Model Size:1.4BQuant:BF16Context Size:32kPublished:Aug 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Ouro-1.4B-Thinking-depth-GRPO by omar81939 is a 1.4 billion parameter causal language model. This model incorporates a 'Thinking-depth' mechanism, allowing for configurable processing steps during inference. It is designed for use cases where controlling the computational depth per token generation is beneficial, offering flexibility in balancing performance and resource usage.

Loading preview...

Ouro-1.4B-Thinking-depth-GRPO Overview

Ouro-1.4B-Thinking-depth-GRPO is a 1.4 billion parameter causal language model developed by omar81939. A key feature of this model is its integration of a 'Thinking-depth' mechanism, which allows users to specify the number of 'unrolled transformer steps' (total_ut_steps) during inference. This parameter can be configured to control the computational depth applied per token generation.

Key Capabilities

  • Configurable Inference Depth: Users can set the total_ut_steps parameter (e.g., DEPTH = 16) to adjust the model's processing depth.
  • Integration with Hugging Face Transformers: The model is designed for straightforward loading and usage with the transformers library.
  • vLLM Compatibility: It supports deployment with vLLM, requiring hf_overrides={\"total_ut_steps\": DEPTH} for engine creation.

Good For

  • Resource-constrained environments: The ability to adjust thinking depth can help optimize performance for specific hardware or latency requirements.
  • Experimental AI research: Researchers can explore the impact of varying computational depth on model outputs and efficiency.
  • Applications requiring flexible inference: Use cases where the trade-off between inference speed and output quality needs fine-tuning based on the task.

Finetunes
1 models