Open-Source

Moonshot AI hat am 16. Juli 2026 ein Modell mit 2,8 Billionen Parametern veröffentlicht. Offene Gewichte. 1 Million Token Kontext. Integrierte Bild- und Videoerkennung. Preis: $3 pro Million Input-Tokens (ca. 2,70 €). Die vollständigen Gewichte (Modified MIT Lizenz) kommen am 27. Juli.

Meine ehrliche Meinung: Das ist der Moment, an dem Open-Source-KI aufgehört hat, die “günstige Alternative” zu sein — und zu einem echten Architektur-Wettbewerb geworden ist. Kimi Delta Attention und Attention Residuals sind kein Marketing-Geschwätz. Das sind echte Engineering-Entscheidungen für Effizienz, nicht nur für rohe Größe. Das ist wichtiger als die Billionen-Parameter-Schlagzeile.

Was mich als Ingenieur begeistert

2,8B Parameter, aber nur ein Teil wird pro Anfrage aktiviert. Das ist clever. Rohle-Skalierung ist eine Sackgasse — Bandbreite und Strom gewinnen am Ende immer. Moonshot hat das verstanden.

Was ich noch nicht kaufe

“Zweitbeste nach den Top-Proprietary-Modellen” ist Moonshots eigener Benchmark, nicht unabhängig. Ausführlichkeit-Probleme und Aufgaben-Fehler tauchen schon in der echten Nutzung auf. Ich kenne diesen Film — großer Launch, große Behauptungen, leise Korrektur drei Wochen später.

Meine Prognose

Dieses Modell wird richtig gut sein für agentisches Coden und lange Kontexte — und für alles andere ziemlich überhypt. Die Veröffentlichung der Gewichte am 27. Juli ist der echte Test. Bis dahin ist es Marketing mit sehr gutem Engineering drunter.

Was das für lokale Inferenz bedeutet

Für alle, die lokal laufen lassen — wie ich auf meinem Mac mini — ist das bei 2,8B Parametern noch außer Reichweite. Aber es bestätigt die Richtung: sparse MoE, effiziente Attention, offene Gewichte. Das ist das Playbook, das kleinere Labs in den nächsten zwei Jahren kopieren werden.

Einen Blick wert. Aber noch nicht den Stack umstellen.

I’ve been following the German government’s slow drift away from Microsoft for a while now, and Mecklenburg-Vorpommern just made a move that caught my attention. They’re not just talking about open source — they’re actually rolling it out.

What’s Happening

The state is replacing Microsoft SharePoint with Nextcloud for its entire administration. We’re talking about more than 50,000 public employees eventually. Around 5,000 are already on it, and the roadmap includes chat, video meetings, and full groupware — essentially rebuilding the Microsoft 365 stack with open-source tools.

OpenProject handles their project management. And here’s the part I really like: they built their own AI assistant called LEA, based on OpenWebUI. No ChatGPT. No Copilot. They’re using European AI models like Mistral and Tilde — keeping everything away from US providers entirely.

Why This Matters

As someone who self-hosts as much as possible, I love this. Data sovereignty isn’t just a buzzword when a government actually acts on it. When your collaboration platform, your AI tools, and your project management all run on infrastructure you control, you own the whole chain.

The officials behind this say they expect greater control, improved security, and long-term cost savings. That tracks. When you’re not paying per-seat licensing to a single vendor, the economics shift. And when you can audit every line of code, security becomes something you verify, not something you trust.

The Bigger Picture

This isn’t happening in isolation. Schleswig-Holstein already pulled off a similar migration successfully, and Mecklenburg-Vorpommern took inspiration from that. Multiple German states are now actively reducing their Microsoft dependency. It’s becoming a trend.

They’re not replacing Windows with Linux on desktops — yet. That’s still a bridge too far for most government IT departments. But the backend infrastructure is shifting, and that’s where the real power lives anyway.

Competition Is Good

I think this is healthy for the whole ecosystem. Having a different provider gives users a different approach to the same problem. Microsoft isn’t going to disappear, but when governments have real alternatives, the pressure to innovate and respect user data goes up for everyone.

Whether you care about open source, data privacy, or just want to see big tech face some real competition — what’s happening in northern Germany is worth watching. This isn’t a fringe experiment. It’s a state government betting that open source can run at scale, and putting 50,000 employees on the line to prove it.

I’m rooting for them.

I’ve been running AI models locally for about a year now, and it completely changed how I think about privacy. Every word I type stays on my machine — no servers, no telemetry, no third parties.

Cloud tools like ChatGPT and Claude are powerful, but many services store your input on remote servers. That trade-off bothered me when working with sensitive code or private documents, so I started exploring local AI.

Why Go Local?

• Absolute Privacy: My prompts never leave my machine. No data goes to any cloud provider, period.
• No Recurring Costs: After the initial hardware investment, running models is free.
• Offline Capability: I’ve used local AI on flights and during internet outages. It just works.
• Full Control: Run any open-source model without rate limits or usage caps.

What I Actually Use It For

This isn’t just theory — I use local AI daily in two practical ways:

Speech-to-text on My Mac: I have a local model on my Mac for speech-to-text. I speak, it writes — that’s how I’m writing this post right now.

News Summarization on My iPhone: I run a small model as an iPhone app. Long articles get fed in, and I get a clean 150-200 word summary. Fully offline.

Model Sizes

The “B” stands for billions of parameters — more is generally smarter, but needs more hardware.

• 7B (Llama 3 8B, Mistral 7B): My go-to. Runs on consumer hardware, handles summarization, coding, and conversation well.
• 13B (Llama 2 13B, CodeLlama 13B): Smarter, especially for complex reasoning and longer code. Needs more VRAM.
• 70B (Llama 2 70B): Approaching cloud-level quality, but requires 48GB+ VRAM or heavy quantization.
• 141B total / 22B active (Mixtral 8x22B): A Mixture-of-Experts model with 8 experts — uses ~22B parameters per forward pass. Requires 48GB+ VRAM or heavy quantization.

VRAM Requirements

• 8GB: Sweet spot for quantized 7B models — what I started with.
• 16GB: Handles 13B models and some quantized 30B models.
• 24GB+: Unlocks larger quantized models and 13B at full precision.
• 48GB+: 70B models become accessible, especially on Apple Silicon with unified memory.

Getting Started

1. Install Ollama: Head to ollama.com and download the installer. Takes about two minutes.
2. Pull your first model: Run ollama run llama3 in a terminal. Downloads the 8B model (~4.7GB) and starts chatting. That’s it.
3. Add a web interface: Run docker run -d -p 3000:8080 ghcr.io/open-webui/open-webui:main and open http://localhost:3000 for a ChatGPT-like experience.
4. Explore: Try ollama pull mistral or ollama pull codellama for different use cases.

Troubleshooting

Running on CPU instead of GPU? On Linux, install NVIDIA drivers first. Out of memory? Try a smaller model or Ollama’s :q4_0 tag for lighter quantization. Everything runs fully offline once downloaded.

Personal Take

Local AI shines for sensitive data, offline work, and experimenting without API costs. That said, I still use cloud AI when I need the latest models or large context windows — it’s not either-or.

Note: Hardware prices and model availability may change. Always check current pricing before purchasing.

Disclosure: Some links in this post may be affiliate links. I may earn a commission at no extra cost to you.

💻 Hardware Recommendation: To run local LLMs smoothly, you’ll need a GPU with at least 16GB of VRAM or an Apple Silicon Mac. See my GPU Buying Guide for recommendations.