TokenRhythm/NeoHorse

What it solves

NeoHorse-1 addresses the challenge of creating models capable of recursive self-improvement (RSI). It aims to move beyond static training by creating a loop where models are evaluated, selected, and updated based on their actual performance in agentic tasks, tool use, and coding.

How it works

NeoHorse-1 uses a "routing harness" that manages a pool of different models. This harness assigns tasks, records how tools are used and the resulting outcomes, and estimates the demand for specific capabilities. This feedback is then fed into the next training mixture. The project employs an agentic post-training framework featuring routing-guided curriculum SFT (Supervised Fine-Tuning) and on-policy distillation, which converts execution trajectories into training signals while maintaining the context of the harness.

Who it’s for

It is designed for developers and researchers focusing on agent-native causal language models, specifically those needing efficient local deployment (4B model) or higher capacity (9B model) for tool-use and instruction-following tasks.

Highlights

• Recursive Self-Improvement Prototype: Implements an evaluation–selection–update loop to iteratively improve model capabilities.
• Agentic Post-Training: Uses routing-guided curriculum SFT and on-policy distillation to refine models.
• High-Quality Data Pipeline: Includes strict duplicate removal, evaluation decontamination, and subscene-level labeling (Scene/Goal/Outcome).
• Flexible Deployment: Available in 4B and 9B parameter sizes with GGUF and quantized versions for easier local hardware execution.