llm.create_chat_completion(
messages = "No input example has been defined for this model task."
)MiniGPT-22M Chat
The smallest known coherent multilingual chatbot at 22.5M parameters.
Model Description
Trained from scratch on Google Colab (A100) by an indie developer. This model is a fully chat-tuned language model that can hold conversations, respond coherently, and operate in 10+ languages — all in an 86MB GGUF file.
Training Details
- Parameters: 22.5M
- Architecture: gpt2
- Training data: SlimPajama + Wikipedia + 3,000 conversation examples
- Total tokens seen: ~4 billion (6.7 epochs over 600M token corpus)
- Overtraining ratio: ~9-10x Chinchilla optimal (intentional for small model)
- Val loss: Slightly below train loss, no overfitting
Performance
- Outperforms Rocket-3B on greeting/conversational behavior
- Comparable to SmolVLM-256M on chat tasks
- 70+ tokens/sec on Pixel 6a
Usage
Load the GGUF in any llama.cpp compatible app such as LM Studio, Ollama, or any other local inference tool that supports GGUF format.
Prompt Format
This model uses a simple User/Assistant format: User: {prompt} Assistant:
- Downloads last month
- 35
Hardware compatibility
Log In to add your hardware
32-bit
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="SmallAICreator/Minigpt-chat", filename="minigpt-rp-sft-f32.gguf", )