Quasar-Tiny

Quasar-Tiny

This is a distilled version of the Quasar Language Model with Parameter Memory Bank and Liquid Neural Networks, featuring true infinite context window capability.

Model Details

• Model Type: Quasar Language Model (Decoder-Only)
• Size: ~200M parameters
• Training: Knowledge distilled from Qwen/Qwen3-0.6B
• Dataset: eyad-silx/Small-QuasarDataset
• Step: 1

Architecture

The Quasar Language Model combines Liquid Neural Networks with a Parameter Memory Bank for unlimited context processing. Key features:

• Decoder-Only Architecture: No positional encoding limitations
• Parameter Memory Bank: Stores and retrieves information from unlimited context
• Liquid Neural Networks: Dynamic state-based processing for better temporal modeling

Usage

from transformers import AutoTokenizer
from quasar_lm_100m import QuasarLM100M
import torch

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("eyad-silx/Quasar-Tiny")
model = QuasarLM100M.from_pretrained("eyad-silx/Quasar-Tiny")

# Generate text
prompt = "The history of artificial intelligence"
prompt_ids = tokenizer.encode(prompt, return_tensors='pt')
output_ids = model.generate(prompt_ids, max_length=100)
generated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(generated_text)