Models to test

Small language models for consumer hardware

Gemma 4 12B for 16 GB RAM Laptops Edit: Tested in another article

Paper, Articles or Videos

• Matamo with Claude access
• I Trained a Close Relative of Neural Networks in Python: Logistic Regression Using NumPy and PyTorch
• Addition is All You Need for Energy-Efficient Language Models
• A clasic Andnrew Karparthy's 1 h Intro to Large Language Models

Series of content

• Sustainable AI
• Civic KI Lernangebote
• S. Willison newsletter about code research
• Helmholtz AI conference 2024 videos
• Practical deep learning
• Deep learning school 2024
• Elements of AI online course—auch auf Deutsch.
• Coding for tomorrow (deutsch)
• Roadmap to learn generative AI in 2025

Paper

• Famouse Attention is all you need
• Making Theft Useless: Adulteration-Based Protection of Proprietary Knowledge Graphs in GraphRAG Systems
• application of large language models for the extraction of information from particle accelerator technical documentation
• Conservative Q-Learning
• Large Language Models for Human-Machine Collaborative Particle Accelerator Tuning through Natural Language by Jan and Annika

Things to test

RAGs

• RAG cheat sheet examples
• GvR slides about RAG
• Tutorial/Blog article, another
• Slides and notebooks of a talk
• How to Create a RAG Evaluation Dataset From Documents

Agents

Agents tutorial Video series by Jetbrains. I wasn't too happy with Pi any my local models yet. Maybe Vibe by Mistral is better, but I fear it's the models' fault.

How Jesse Vincent used coding agents and skills in 2025 and Simon Willison's way of doing so

Research driven agents is a concept originally by Andrew Karparthy to let your agents read research papers before doing the concepts and coding tasks. As my local agents tasks weren't successful yet, I haven't tested this on my own so far.

Others

Quite some time, I haven't heard about new inovations like Striped Hyena once was. I'm wondering if testing it today would be interesting as it is already from 2023, but also “just” 7B parameters large. Also from 2023, Mixture of Experts survived—SwitchHeads paper on using MoE also on the attention layer.

Misc

nVidia RAPIDS (documentation, featured notebooks) is NVIDIA's CUDA stack for (conventional) data science (cuDF instead of Polars/Pandas, cuML instead of scikit-learn, cuGraph instead of NetworkX, etc.)

awesome MLOps list

Practitioner's guide to Reinforcement Learning (RL)

General links

The Trillion parameter consortium is a group of scientists that wants to have a more open, fair LLM models for science.

For comparison of models, the website artificialanalysis.ai is very recommended.