Reasoning Models, Local Agents, and Whatever Just Dropped with Sebastian Raschka
GPT-5.6 and GLM 5.2 look exciting today. By the time Hugo and Sebastian go live, they may be old news. Good. Sebastian's work is built for what survives the release cycle: how models reason, how to test them, and what you learn when you build the pieces yourself.
His new 440-page book, Build a Reasoning Model (From Scratch), starts with a small pretrained Qwen3 model and adds the pieces people usually wave their hands at: evaluation, inference-time scaling, reinforcement learning, and distillation. You can inspect the code, run the experiments, and see where the reasoning behavior actually comes from.
The last time Sebastian joined Vanishing Gradients, he and Hugo took apart hybrid attention, KV caches, post-training, inference scaling, and the questionable mental health effects of supervising agents that supervise themselves. Since then, the book has shipped and local open-weight coding agents have become surprisingly capable.
Sebastian has been running Qwen3.6 through Qwen-Code, Codex, and other coding harnesses. Hugo currently has Qwen3.6 35B-A3B living inside Pi in Ghostty. Naturally, they have notes.
Fresh releases will still get their turn. When labs keep the architecture and training details proprietary, perceived performance may be most of what users can inspect. Hugo and Sebastian will compare notes on whatever just shipped, then get into the parts they can actually test: reasoning behavior, evals, local agents, and the harness around the model.
What We'll Cover
• What you learn about "reasoning" by building the whole path from a base model to evaluation, inference scaling, reinforcement learning, and distillation
• How to test a reasoning model before trying to improve it, including what parsers, verifiers, and benchmark-style checks can actually tell you
• When extra test-time compute earns its keep: choosing between best-of-N, majority voting, and self-refinement, and spotting when more tokens merely produce a more expensive wrong answer
• How reinforcement learning changes a model's reasoning behavior, from verifiable rewards and GRPO-style training to the engineering details hidden by a benchmark score
• What a 22 GB local coding model can actually do on a well-equipped Mac, and which workloads still belong on larger hardware
• How to choose a harness for an open-weight model, and why the same Qwen model can behave differently across Qwen-Code, Codex, Claude Code, and Pi
• Which recent reasoning ideas survived contact with Sebastian's experiments, and which ones looked better on paper
• GPT-5.6, GLM 5.2, and whichever model everyone started arguing about fifteen minutes before we went live
Come for the reasoning models. Stay for two friends comparing local-agent stacks and trying not to get distracted by a model released during breakfast.
Register to join live or get the recording afterwards.
About Sebastian
Sebastian Raschka is an LLM research engineer, author, and educator focused on large language models, reasoning models, deep learning, and practical machine learning systems. He writes Ahead of AI and is the author of Build a Reasoning Model (From Scratch) and Build a Large Language Model (From Scratch).