ScaleLLM AI

scalellm-ai is a tiny high-level Python toolkit for building AI applications with very little code.

It now supports scratch-trained models for:

• LLMs
• image generation
• video generation
• text-to-speech

It also still includes easy wrappers around strong pretrained models for images, video, TTS, vision, speech recognition, embeddings, text generation, and zero-shot classification.

Important reality check

The scratch-trained image, video, and TTS models are real trainable models from random weights, but they are toy-sized starter architectures meant for learning and experimentation.

That means:

• you can genuinely train them from scratch
• they can learn small custom datasets
• they will not match Stable Diffusion, Sora, or commercial TTS quality
• they are designed so the user-facing code stays under 20 lines

Install

Core package:

pip install scalellm-ai

Feature installs:

pip install "scalellm-ai[image]"
pip install "scalellm-ai[video]"
pip install "scalellm-ai[tts]"
pip install "scalellm-ai[game]"
pip install "scalellm-ai[all]"

1. Train your own tiny LLM from scratch

from scalellm_ai import ScaleLLM

llm = ScaleLLM(d_model=128, n_layer=4, n_head=4, max_seq_len=128)
text = open("dataset.txt", encoding="utf-8").read()
llm.train(text, steps=500, batch_size=4)
print(llm.generate("Python is", max_new_tokens=80))
llm.save("my_llm.pt")

2. Train an image model from scratch

Manifest format:

images/cat1.png|orange cat
images/car1.png|red sports car

from scalellm_ai import ImageAI

ai = ImageAI.from_scratch(image_size=32)
ai.train("image_manifest.txt", epochs=20)
ai.generate("red sports car", output="car.png")
ai.save("my_image_ai.pt")

3. Train a video model from scratch

Manifest format:

clips/fly.gif|bird flying
clips/wave.gif|ocean waves

from scalellm_ai import VideoAI

video = VideoAI.from_scratch(image_size=32, num_frames=8)
video.train("video_manifest.txt", epochs=20)
video.generate("bird flying", output="bird.gif", fps=6)
video.save("my_video_ai.pt")

4. Train a TTS model from scratch

Manifest format:

hello.wav|hello world
welcome.wav|welcome to scale llm ai

from scalellm_ai import TTS

voice = TTS.from_scratch(sample_rate=8000, seconds=1.0)
voice.train("tts_manifest.txt", epochs=20)
voice.speak("hello world", output="hello.wav")
voice.save("my_tts.pt")

5. Pretrained image generation

from scalellm_ai import ImageAI

ai = ImageAI()
ai.generate("cinematic robot exploring an ancient library", output="robot.png")

6. Pretrained video generation

from scalellm_ai import VideoAI

video = VideoAI()
video.generate("a small spacecraft flying through glowing blue clouds", output="space.mp4")

7. Pretrained TTS

from scalellm_ai import TTS

voice = TTS()
voice.speak("ScaleLLM AI can turn this sentence into speech.", output="speech.wav")

8. Game-playing AI

from scalellm_ai import GameAI

agent = GameAI("CartPole-v1", algorithm="PPO", verbose=1)
agent.train(20_000)
agent.save("cartpole_agent")

9. Vision AI

from scalellm_ai import VisionAI

vision = VisionAI()
for result in vision.classify("photo.jpg", top_k=3):
    print(result["label"], result["score"])

10. Speech-to-text

from scalellm_ai import SpeechAI

speech = SpeechAI()
print(speech.transcribe("meeting.wav"))

11. Embeddings / semantic similarity

from scalellm_ai import EmbedAI

embed = EmbedAI()
print(embed.similarity("A dog is running outside.", "A puppy is playing outdoors."))

12. Pretrained text generation

from scalellm_ai import TextAI

ai = TextAI()
print(ai.generate("Explain neural networks simply:", max_new_tokens=120))

13. Zero-shot classifier

from scalellm_ai import ClassifyAI

ai = ClassifyAI()
result = ai.classify("The team won the championship last night.", ["sports", "business", "science"])
print(result["labels"][0])

14. Music / text-to-audio

from scalellm_ai import MusicAI

music = MusicAI()
music.generate("energetic retro arcade synthwave with a heroic melody", output="theme.wav")

15. Object detection

from scalellm_ai import ObjectAI

ai = ObjectAI()
for obj in ai.detect("street.jpg", threshold=0.7):
    print(obj["label"], obj["score"], obj["box"])

Notes on scratch datasets

• ImageAI.from_scratch expects a manifest with image_path|caption
• VideoAI.from_scratch expects a manifest with video_path|caption
• TTS.from_scratch expects a manifest with audio_path|text
• all scratch models start from random weights
• all scratch models are kept intentionally small so they can be understood and extended

Examples

Every file in examples/ is intentionally 20 lines or fewer.

Publishing safely

Install the publishing tools first with pip install "scalellm-ai[dev]".

Never put PyPI tokens in publish.py. Set the token in your shell instead:

PowerShell:

$env:PYPI_API_TOKEN="pypi-..."
python publish.py

macOS/Linux:

export PYPI_API_TOKEN="pypi-..."
python publish.py

License

MIT. Individual pretrained models can have their own licenses and usage terms; check model cards before redistribution or commercial use.