Eric Schreiber

Hi, I'm Eric, working in research at Aleph Alpha on large-scale neural networks .

My work focuses on efficiency at scale, designing models and systems that are both numerically robust and fast to train and run. I'm interested in neural network architecture, optimization, and high-performance parallel code.

GitHub  /  Google Scholar  /  X  /  LinkedIn

profile photo

Publications

project image

Kolibri: A Sovereign European Model on the Pareto Frontier


Aleph Alpha, et al.
arXiv, October 2026
arxiv / huggingface /

Kolibri is a 78B-A3B open model, trained on 24T tokens with over 20% German data and sits on the Pareto frontier of quality versus serving cost in both English and German.

project image

Reasoning Language Models: A Blueprint


Maciej Besta, Julia Barth, Eric Schreiber, Ales Kubicek, Afonso Catarino, et al.
arXiv, January 2025
arxiv / code /

A blueprint for reasoning language models (RLMs), showing building blocks for structured reasoning, search, and RL-based rollouts. Implementing x1, a composable RLM framework.

project image

CheckEmbed: Effective Verification of LLM Solutions to Open-Ended Tasks


Maciej Besta, Lorenzo Paleari, Marcin Copik, Robert Gerstenberger, Ales Kubicek, et al.
arXiv, June 2024
arxiv / code /

Using embedding vectors of multiple solutions to a given task to gauge their similarity and infer from this the likelihood of a solution being correct.



Blogposts

project image

Inference Performance from First Principles


blogpost /

Prefill is compute-bound, decode is memory-bound, and MoEs make everything more interesting: modelling LLM inference speed from compute, memory, and communication.

project image

Demystifying CuTe: Building High-Performance GEMMs


blogpost / code /

Part 2 of 2 — From copy kernels to fast matrix multiplication

project image

Demystifying CuTe: Understanding Layout Algebra


blogpost / code /

Part 1 of 2 — From conceptual foundations to efficient memory operations

project image

MoE Inference Economics from First Principles


blogpost /

DeepSeek 🐳, Kimi, synthetic data markets and the token overcapacity issue

Projects


Design and source code from Jon Barron's website