Think Twice: The Metis-1.6 Technical Report

Think Twice. The Metis-1.6 Technical Report. By Giulianno Vollmer, from Lernex Research. Published on the twenty-seventh of October.

Metis starts from a bet. Better data, better training methods and new architectures, it wagers, can buy what a bigger parameter count cannot, and at a fraction of the cost. A bet of that kind is only settled by building something and measuring both sides of it: what the model scores, and what it took to get there.

Metis-1.6 is the latest step: two language models built from scratch by a single researcher. Praxis has two point nine four billion parameters, with zero point four six billion active per pass. Logos has ten point two three billion parameters, with one point one eight billion active per pass. The title's think twice is literal. Both models run a hybrid Mamba-2 and attention mixture-of-experts stack twice for every token, so each token gets a second look.

Training came in three phases for each model. Pretraining took eight hundred and four point seven six billion tokens. Context extension took eighteen billion tokens, stretching the context to one hundred and thirty-one thousand and seventy-two tokens. Supervised fine-tuning took a further two hundred billion tokens. Set against the cost side of the bet, that is the budget the scores that follow have to be read against: a training run measured in hundreds of billions of tokens, carried out by one person.

The report is plain that Metis-1.6 was never meant to challenge frontier models, and across forty post-trained benchmarks it mostly does not. Logos averages fifty-two point nine three, which places it seventh of seventeen models in its comparison set and within a point of Qwen3-4B. Its case is efficiency. It reaches that score on roughly a twelfth of the median training data of the comparators that disclose theirs, and with fewer active parameters per token than any model that beats it. That puts it on the Pareto frontier for both.

Mathematics is the strong suit. There it ranks second of thirty models. On AIME two thousand twenty-five it scores sixty-seven point five percent. On HMMT February two thousand twenty-six it scores sixty-three point six four percent. The weak spots are named just as directly: the models are weak at code, at instruction following, and at long context. A model built to compete on cost, in other words, does not claim to have bought everything with it.

By its author's account, the most valuable part of the report is what went wrong. The optimizer that was configured never ran, so attention trained with AdamW for all of pretraining. A one point seven seven billion parameter n-gram memory learned to switch itself off. Five separate fine-tuning choices cost capability.

The subtlest failure was a length reward under group-mean normalization. Once a group's pass rate passed two thirds, the reward punished correct solutions for being long. The author proves that sign inversion and fixes it with a one-line clamp. The report ends on the ten billion token ablations behind MoRE, a Mixture of Recursive Experts, which is the looped architecture of Metis-1.7. A technical report usually lists what worked. This one spends as much care on the mistakes, and treats them as findings in their own right.