Aurora-One-Mini

Aurora One Mini — 124M

Aurora One Mini is a compact, community-built language model designed for fast local chat, experiments, and lightweight AI applications.

At only 124 million parameters, it is small enough to run comfortably on ordinary laptops and edge devices while remaining useful for short-form generation and experimentation.

What makes it interesting

• Tiny and fast: practical for local inference and rapid prototyping
• Native ChatML format: structured user/assistant conversations
• Hugging Face + GGUF exports: works with Transformers and llama.cpp-compatible tools
• Open experiment: trained and evaluated on a single consumer GPU

Model details

• Architecture: GPT-style causal language model
• Parameters: approximately 124M
• Layers: 12
• Hidden size: 768
• Attention heads: 12
• Context length: 1,024 tokens
• Vocabulary: GPT-2 BPE plus ChatML control tokens
• Final pretraining: 45,000 steps, approximately 15 tokens per parameter
• Released checkpoint: deterministic v2, step 2,000 of targeted post-training

Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "North-ML1/Aurora-One-Mini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "What is the capital of France?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=80,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

GGUF files

The companion GGUF files are provided for local runtimes:

• aurora_one_mini_deterministic_v2_f16.gguf — highest fidelity
• aurora_one_mini_deterministic_v2_q4_k_m.gguf — compact CPU-friendly quantization

Use the Q4_K_M file for a fast, low-memory demo. Use the F16 file when preserving maximum quality is more important.

Honest limitations

This is an experimental 124M model, not a frontier assistant. It can produce fluent short responses, but it may hallucinate, repeat itself, or answer arithmetic and factual questions incorrectly. For dependable applications, pair it with a calculator, retrieval system, memory layer, and explicit output validation.

The native-ChatML factual smoke test scored 3/20 on a small internal suite. This score is reported to set realistic expectations and should not be interpreted as a general benchmark.

Intended use

Good fits include:

• local chat experiments
• educational model training projects
• embedded or low-resource inference
• prompt-format and agent-runtime experiments
• fast prototyping with Transformers or llama.cpp

Avoid using it as the sole source of truth for medical, legal, financial, safety-critical, or factual decision-making.

Prompt format

The model was post-trained using ChatML-style turns:

<|im_start|><|user|>Your question<|im_end|>
<|im_start|><|assistant|>

The included tokenizer metadata contains the required special tokens.

Acknowledgements

Aurora One Mini was trained as a small-scale independent experiment using PyTorch and a consumer NVIDIA GPU. Contributions, evaluations, and improvements are welcome.

License

Released for research and experimentation. Add the project’s final license here before redistributing commercially.