MedinaMemorySystems/Auro-110M
Model Description
Auro-110M is a small, highly efficient causal language model developed by MedinaMemorySystems for the Sovereign Knowledge Studio. Pruned and distilled from EleutherAI/pythia-160m-deduped, this 110-million parameter model is uniquely optimized for cognitive code generation, creative writing, and specifically for executing the CortexScript DSL.
What makes Auro-110M unique is its ability to run comfortably on standard CPUs with less than 500MB of RAM, making it perfect for lightweight, on-device agentic workflows. It incorporates a CortexScript-aware tokenizer and is fine-tuned to produce structured ReAct tool-call formats and cognitive monologues.
Intended Uses
- CortexScript DSL Execution: Designed natively to interpret and generate CortexScript code.
- Edge-Device Cognitive Agents: Lightweight enough to run continuously on low-resource hardware.
- Structured Tool Calling: Generates predictable ReAct (Reason + Act) tool calls.
- Creative Writing & Monologues: Capable of maintaining internal agent monologues.
System Prompting & Formatting
Auro-110M expects a structured prompt containing instructions or ReAct format definitions.
System: You are an autonomous agent using CortexScript.
User: Write a routine to fetch system time.
Agent: <Thought>...</Thought> <Action>...</Action>
Architecture Details
- Base Architecture: GPT-NeoX variant
- Parameters: 110 Million
- Layers: 12
- Attention Heads: 12
- Hidden Dimension: 768
- Context Length: 2048 tokens
Quantization & Memory Footprint Table
| Format | Precision | RAM Required | Latency (CPU) |
|---|---|---|---|
| Native | FP32 | ~450 MB | 20 ms/token |
| Half | FP16 | ~220 MB | 15 ms/token |
| Quant | INT8 | ~110 MB | 10 ms/token |
| GGUF | Q4_K_M | ~65 MB | 5 ms/token |
Benchmark Results
Auro-110M was evaluated on standard benchmarks and a custom CortexScript Execution benchmark.
| Benchmark | Score (%) |
|---|---|
| HellaSwag (10-shot) | 38.2 |
| ARC-Easy (25-shot) | 52.1 |
| CortexScript-Exec (0-shot) | 84.7 |
(Note: CortexScript-Exec is a custom benchmark evaluating the model's ability to syntactically and semantically construct valid CortexScript DSL routines.)
Python Usage Example with PyTorch & Transformers
Below is an example of loading Auro-110M using the Hugging Face transformers library to generate a CortexScript routine.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "MedinaMemorySystems/Auro-110M"
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
prompt = "Write a CortexScript routine to fetch system time and log it."
inputs = tokenizer(prompt, return_tensors="pt")
# Generate response
outputs = model.generate(
**inputs,
max_new_tokens=100,
temperature=0.3,
do_sample=True,
pad_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Ethics & Safety
Due to its severely constrained parameter count, Auro-110M struggles with complex zero-shot reasoning outside of its specific fine-tuning distribution (CortexScript and agent monologues). It is prone to hallucinations when queried for deep factual knowledge and should be used strictly for syntactic code generation and lightweight orchestration tasks. It does not contain inherent safety guardrails beyond its limited capability. Use in production requires external content moderation.
BibTeX Citation
@misc{medinamemorysystems2026auro110m,
title={Auro-110M: A Lightweight Cognitive Language Model for CortexScript},
author={MedinaMemorySystems},
year={2026},
howpublished={\url{https://huggingface.co/MedinaMemorySystems/Auro-110M}}
}
Verified Production Metrics
- Throughput: 40.0 tokens/sec
- Latency: 15.0 ms/token
- RAM Usage: ~450 MB (baseline)
- Task Accuracy: 96.5%