Energy Consumption: ~100 kWh of electricity (enough to run a normal refrigerator for a month).
Carbon Footprint: Roughly 40 kg of CO₂ pumped directly into the atmosphere.
Water Evaporated: ~200 liters of freshwater boiled off at power plants to cool the grid.
Collateral Damage: Statistically speaking, 0.0004 innocent people probably developed a mild respiratory cough from these emissions. at least one local houseplant died because the resident cat got scorched by the GPU exhaust, ricocheted around the room in an absolute panic, and violently knocked the pot onto the floor.

🎭 Generation Examples

Do not use this model in production. Do not use it for research. It generates pure, unadulterated slop.

Prompt: 'The capital of France is'

Generated: 'the capital.'

Prompt: 'The capital of France is ' (Notice the trailing space)

Generated: '12.089 - 37,6407 -2052'

Prompt: 'syndrome WeOneGuy this is '

Generated: '10.2%\n- The most common way of developing a neurological disease is to take a group of people with a neuronous disease that is not always a problem in the development of neurodegenerative diseases and they may also have a neural disorder.\n- In'

Prompt: 'In 1969, NASA'

Generated: '’s Space Agency (HOLA) launched a 1967 spaceflight campaign in the Alaska.''

🔬 Architecture & Paper Density

The model uses exactly 38.47M parameters. Output projections are tied directly to the embeddings (weight tying) because we needed to save space.

Let's look at the Maximum Paper Density:

$$ \frac{20 \text{ papers}}{38,470,000 \text{ parameters}} \approx 5.2 \times 10^{-7} \text{ papers/parameter} $$

If we flip the math, that is 1.92 million parameters per academic paper. If we pack any more cutting-edge AI research into a model this small, the weights will undergo spontaneous nuclear fusion.

💻 Usage

If you absolutely must run this for some morbid reason:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Minarut/ANA-LM-38M"

tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True).to("cuda")

inputs = tokenizer("The capital of France is ", return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    do_sample=True,
    temperature=0.8,
    top_p=0.9,
    repetition_penalty=1.1,
)

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

> The capital of France is 29% less efficient than the rest of Atlantis.