AresXLG1: An ultra-compact 100M parameter language model
2 min readupdated
Developed by the ArrowLabs team, AresXLG1 is part of the Ares family and delivers strong results on the Open SLM Leaderboard (the top benchmark for compact models). It was trained on 50 billion tokens (text fragments for analysis) using carefully filtered datasets, including DCLM, UltraFineWeb L3, and industrial data arrays. The architecture and tokenizer are inspired by the latest KLM models, while AresXLG1 outperforms them across several metrics. As a base model without dialogue fine-tuning, it is ideal for researchers and developers building custom fine-tunes.
Key features
- 100 million parameters for fast performance on lightweight hardware
- Trained on 50 billion tokens from verified sources
- Strong benchmark results (ArithMark 3.0, PIQA, HellaSwag, ARC)
- Architecture inspired by the latest KLM models
What you can do
- Use as a base model for task-specific fine-tuning
- Research the capabilities of ultra-compact language models (SLMs)
- Integrate into local text processing pipelines with minimal resource requirements
Example
The model is a base checkpoint without instruction tuning. It is not designed for direct dialogue or out-of-the-box question answering, and is instead used for fine-tuning or text continuation.
Read more
DetailsBenchmarks: ArithMark 3.0 — 60.60, PIQA — 68.25, HellaSwag — 46.58, ARC-Easy — 58.00, ARC-Challenge — 32.08, Open SLM Int Index — 34.84
Hardware: trained on 4x B300 GPUs
Data volume: 50 billion tokens
Limitations: this is a base model without instruction or reasoning tuning, and it cannot maintain a conversation or follow instructions










Comments