Metis-1.6 models reach math benchmarks on low data
Original title: Think Twice: The Metis-1.6 Technical Report
Praxis and Logos achieve strong math scores with hybrid Mamba-2 architecture using fraction of training data of comparators.
Metis-1.6 releases two models built from scratch: Praxis with 2.94B parameters and Logos with 10.23B. Both use hybrid Mamba-2 and attention MoE stacks trained on 1.02T tokens total. Logos ranks seventh in general benchmarks but second in mathematics among 30 models. The report details training choices that underperformed and ablations for the upcoming MoE architecture.
- Praxis: 2.94B parameters, 0.46B active
- Logos: 10.23B parameters, 1.18B active
- Trained on 1.02T tokens
- Strong on AIME 2025 and HMMT February 2026
- Weak at code and instruction following
From the original article
1 min read at huggingface.co
Metis is a bet that better data, better training methods and new architectures can buy what a bigger parameter count cannot, at a fraction of the cost. Metis-1. 6 is its latest step: two language models built from scratch by a single researcher, Praxis (2.
By Giulianno Vollmer · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.