TARS-1B

TARS-1B is a 1 billion parameter Liquid Neural Network language model built from scratch without using transformers or attention mechanisms. The architecture is designed for continuous-time reasoning and efficient generalization.

TARS:

• Architecture: Non-transformer, non-attention LNN (Liquid Neural Network)
• Training Tokens: Only 300 million tokens (Still training)

---

📊 Benchmarks

Benchmark	Accuracy
MMLU-Redux (Math)	25.4%
ARC-Easy	32.95%
PIQA	51.36%
BoolQ	66.00%

⚠️ These results were achieved with only 300M tokens of pretraining, making them exceptionally efficient compared to transformer-based models.

---

🚧 Training Status

We are still actively pretraining the model to reach higher performance on reasoning, coding, and complex tasks. Expect significant improvements as we scale toward billions of tokens.

---

🧪 Research Direction

TARS-1B is part of a broader research effort at SILX AI to build language models that generalize better with minimal compute using biologically inspired architectures.

---