A lightning‑fast, open‑source AI assistant forged in the heart of the TinyModels community. Hand‑picked data, real‑world training, and a personality that doesn't feel like a corporate robot. This is our take on what a compact, genuinely useful model should be.

⬇ GET STARTED

360M

Parameters

724 MB

Size (FP16)

1 Epoch

Focused Training

Apache 2.0

License

🚀 QUICK START

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "TinyModels/Atom-350M",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("TinyModels/Atom-350M")

messages = [
    {"role": "user", "content": "Explain how a bicycle stays upright in simple terms."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

🔧 OUR RECIPE

BASE MODEL

SmolLM2‑360M‑Instruct

FINE‑TUNING

QLoRA (4‑bit, rank‑16)

TRAINING DATA

SmolTalk (smol‑magpie‑ultra)

FRAMEWORK

Unsloth + Hugging Face

HARDWARE

Kaggle T4 (free tier!)

EPOCHS

1 — lean & efficient

🤍 THIS IS OURS

We didn't just download a model and slap a name on it. TinyModels hand‑picked the training data, configured the QLoRA adapters, and ran the entire training pipeline ourselves. Atom‑350M is our interpretation of a small, helpful, open assistant — built with pride, released with no strings attached.

If you do something cool with it, just give us a shout‑out. We'd love to see what you build.

📜 LICENSE

Apache 2.0 — free for research, commercial use, and even intergalactic travel.

Crafted by TinyModels — small models, big ambitions.