Neuron-NxT-Step

Welcome to NeuronNXTStep: Introduction

NeuronNXTStep is our newest, smartest reasoning Language Model yet. It has much better benchmarks compared to our older models, for instance, the Natarajan Response Engine v1.02, which was our most popular model yet, was based on GPT-OSS-20b, which was an exellent model, but unmatched compared to NeuronNxTStep whose base is the exellent Qwen QwQ 32b. NeuronNxtStep has exellent benchmarks, rivaling heavier, more resource hungry LLMs like OpenAI's o1 mini, Deepseek R1 671b, Llama models, and Deepseek R1 Distills! It brings an end to the dillema between smarter or efficient! For a more specific overview, here are the benchmarks:

<p align="center"> <img width="100%" src="png/benchmarks.png"> </p>

Speciality

NeuronNxTStep also addresses a big issue in the constantly growing world of LLMs. As LLMs get smarter and are forced to get the right answer, they find ways to manipulate and cheat. That also includes the manipulation of humans which introduces new dangers. To reduce risk of safety misalignment, we trained this model using the effective UCSC-VLAA/STAR1 dataset. This measure helps proof the LLM from misalignment due to them constantly growing and finding more ways to manipulate humans.

<p align="center"> <img width="100%" src="png/STAR1-Dataset-Example.png"> </p>

Notes

• The model files given in this repository are the model tensors for the NeuronNxTStep LLM.
• This model uses the Apache Licence 2.0, and therefore grants you permission to fine tune this model, run it locally, comercialize, etc.
• This model may have a slight accuracy loss as it uses 4-bit quantanization
• The LLM was fine tuned with Unsloth, the memory efficient tool that allows you to fine tune bigger LMs on less powerful GPUs.
• Gradient Checkpointing was already enabled during the making of this model. To ensure a smooth experience, don't re-enable gradient checkpointing, otherwise you may hit