# Metis-1.6 report details two small language models and their failures

> Researcher Giulianno Vollmer reports that Metis-1.6’s Logos model scored well in math benchmarks but struggled with code, instruction following and long context.

Oossa · 2026-10-30 · https://oossa.com/en/metis-1-6-report-details-two-small-language-models-and-their-failures

Giulianno Vollmer’s Metis-1.6 technical report describes two language models built from scratch: Praxis, with 2.94 billion parameters, and Logos, with 10.23 billion. The report says Logos placed seventh among 17 models across 40 benchmarks, and second among 30 models in mathematics; it also notes weaknesses in code, instruction following and long-context tasks.

Vollmer details several training problems, including an optimizer that never ran and a length-reward bug that could penalize correct, lengthy answers. The report says Logos was trained on 1.02 trillion tokens and was not intended to compete with frontier models.

## The facts

- Logos scored 67.50% on AIME 2025 and 63.64% on HMMT February 2026, according to the report.
- The report says Metis-1.6’s two models were trained on 1.02 trillion tokens each.

## Why it matters

For developers considering smaller models, the report offers benchmark results and specific training lessons, but its own findings show Logos is weak at code and long-context work; independent validation is not documented here.

## Sources & references

1. [Think Twice: The Metis-1.6 Technical Report](https://arxiv.org/abs/2610.37903) – arXiv cs.LG, 2026-10-30

Last updated: 2026-10-30