Soverius AI

Dear there ,

Three things this week: a free live webinar on coding with local models, Murat's new deep dive on building a local AI coding agent from scratch, and the launch of our new sister newsletter, Quantized AI News.

🎙️ Live Webinar: Software Development with Local LLMs

Tuesday, June 30, 2026, 4:00 PM — free & live

Build smarter. Code locally. Stay in control.

It is time. Token costs rise constantly — so are local LLMs ready to take over? Join our free webinar and see for yourself.

In part one, Rainer shows the setup of a local LLM and connects it to a modern harness like Claude Code. In part two, Murat covers what you need to choose wisely: model types, quantization, the required tools, and hardware requirements.

You'll leave with a ready-to-use setup you can apply at work immediately.

👉 Reserve your seat

🛠️ New Article: Build a Local AI Coding Agent from Scratch

Ever wondered what's actually inside tools like Claude Code or Cursor? In his latest article, Murat strips away the magic: an "AI agent" is just a model plus a harness — a loop with tools wired to the real world. So he builds one from scratch.

The model runs entirely on a laptop: Gemma 4 E4B, quantized to Q8_0, served by llama-server. Along the way you'll learn how to read a model card, what a GGUF file contains, what Q8 quantization really means (with the math), and how to wire tools into the agent loop — until a 4B local model takes "make me a landing page" and writes a real file to disk.

Then comes the uncomfortable part: what happens when you give a model a shell. Murat walks through prompt injection, the "lethal trifecta", and how NVIDIA OpenShell contains the agent with Landlock, an unprivileged user, and a single routed door to the model.

The complete source code is on GitHub, ready to run.

👉 Read the article

📰 Introducing: Quantized AI News

AI moves fast — too fast for anyone to follow the full firehose. That's why we're starting Quantized AI News: a weekly review that compresses the week in AI down to the developments that actually matter, with the same practical, developer-first lens you know from us. Maximum signal, minimal noise — that's the "quantized" part.

The first issue is already out on LinkedIn, and in the coming days Quantized AI News will get its own home on our website as a separate publication.

👉 Read the first issue on LinkedIn

That's it for this issue. See you at the webinar on June 30!

— The Soverius AI Team

Practical AI training for developers and teams

Soverius AI · Austria

Legal notice (Impressum) · Privacy Policy

hello@soverius.ai

View in browser · Unsubscribe · Email preferences