Solo researcher releases Metis-1.6 AI models hitting 67.5% math benchmark accuracy
Tool · Hugging Face · stat: 67.50% A solo researcher published technical details for Metis-1.6, introducing two language models built on a hybrid Mamba-2 mixture-of-experts architecture. The larger…
Tool · Hugging Face · stat: 67.50%
A solo researcher published technical details for Metis-1.6, introducing two language models built on a hybrid Mamba-2 mixture-of-experts architecture. The larger Logos model scored 67.50% on AIME 2025 math benchmarks while utilizing roughly one-twelfth the median training data of comparable systems, totaling 1.02T tokens.
Hybrid Mamba architectures prove solo builders can compete on specialized math tasks Architectural efficiency lets bootstrap teams train targeted reasoning models at a fraction of traditional pre-training compute budgets.
Hugging Face technical paper
Per technical documentation published on Hugging Face.
Supervised fine-tuning comprised nearly a fifth of total training tokens, an unusually high share for open language models.
Every claim ties to a primary source. See our methodology.