NeoHorse-1-4B
An open-source small model with approximately 4 billion parameters developed by the Token Rhythm team. Built on the Coin 3.5 framework, it supports VLM multimodality and employs a recursive self-improvement training paradigm, enhancing capabilities by observing real-time routing system decision logs and through curriculum learning.
Timeline (last 90 days)
NeoHorse-1-4B is an open-source model with approximately 4 billion parameters, based on the Coin 3.5 framework, developed by the Token Rhythm team.
NeoHorse-1-4B improves itself by observing the decision-making process of a real-time routing system, rather than by learning from static samples.
NeoHorse's training data is fed into the model in order of difficulty, accumulating capabilities in a progression from easy to hard.
In the second round of training, NeoHorse has a more powerful teacher model observe and correct its attempts in real time, rather than directly providing pre-written standard answers.
NeoHorse's self-recursive loop occurs during the training phase and is not triggered during the inference phase.
In the blogger's GlacierWatch glacier flood warning system demo, NeoHorse-1-4B ran through the Hermes agent, autonomously discovered and fixed a logical bug where the system displayed 'no personnel in the inundation zone,' and after the fix, correctly displayed that more than 41,000 people were in the inundation zone.
The model exhibited language drift during the recursive reasoning process, with content suspected to be Russian appearing mixed into the thinking process.
The development team compared NeoHorse-1-4B against five small models of approximately 4 billion parameters, and the model ranked among the top in categories such as agent tasks, programming, and instruction execution, with balanced average scores across 10 benchmark tests.
NeoHorse-1-4B exhibited unstable character identity and confused references in creative writing tasks, and the blogger considers its creative writing capability inferior to the original Coin model.
Recursive training makes extensive use of task completion rates and code execution results as reward signals, lacking explicit supervision for narrative consistency, which may lead to catastrophic forgetting of general language modeling capabilities, and this is more pronounced in models with small parameter counts.
1 more timeline events
All Facts (11)
NeoHorse-1-4B is an open-source model with approximately 4 billion parameters, based on the Coin 3.5 framework, developed by the Token Rhythm team.
50%UnverifiedNeoHorse-1-4B improves itself by observing the decision-making process of a real-time routing system, rather than by learning from static samples.
50%UnverifiedNeoHorse's training data is fed into the model in order of difficulty, accumulating capabilities in a progression from easy to hard.
50%UnverifiedIn the second round of training, NeoHorse has a more powerful teacher model observe and correct its attempts in real time, rather than directly providing pre-written standard answers.
50%UnverifiedNeoHorse's self-recursive loop occurs during the training phase and is not triggered during the inference phase.
50%UnverifiedRecursive training makes extensive use of task completion rates and code execution results as reward signals, lacking explicit supervision for narrative consistency, which may lead to catastrophic forgetting of general language modeling capabilities, and this is more pronounced in models with small parameter counts.
50%UnverifiedNeoHorse-1-4B is suitable for local deployment and agentic tasks (programming and bug troubleshooting), but not for creative writing or scenarios requiring long-document consistency.
50%UnverifiedThe development team compared NeoHorse-1-4B against five small models of approximately 4 billion parameters, and the model ranked among the top in categories such as agent tasks, programming, and instruction execution, with balanced average scores across 10 benchmark tests.
50%UnverifiedNeoHorse-1-4B exhibited unstable character identity and confused references in creative writing tasks, and the blogger considers its creative writing capability inferior to the original Coin model.
50%UnverifiedIn the blogger's GlacierWatch glacier flood warning system demo, NeoHorse-1-4B ran through the Hermes agent, autonomously discovered and fixed a logical bug where the system displayed 'no personnel in the inundation zone,' and after the fix, correctly displayed that more than 41,000 people were in the inundation zone.
50%UnverifiedThe model exhibited language drift during the recursive reasoning process, with content suspected to be Russian appearing mixed into the thinking process.
50%