Odysseus System Requirements: What You Actually Need

The most common question about Odysseus, PewDiePie’s self-hosted AI workspace, is not “is it good.” It is “will it run on my machine.” Fair question, because the README’s honest answer is “it depends”: the app itself is lightweight, but local model serving is the heavy part, and that depends on model size, runtime, GPU, and VRAM.

That sentence is true and also useless if you do not already speak VRAM. So here is the translation. This applies to Odysseus, to plain Ollama, and to any local AI setup in June 2026.

the one rule that matters

A local model gets loaded into memory. So the model’s download size has to fit in your RAM (or your GPU’s VRAM) with room left over for the operating system, your browser, and the app itself. A 17GB model on a 16GB laptop is not “slow,” it is a brick.

Everything below follows from that one rule. Model sizes are from the live Ollama library pages for qwen3.5, gemma4, and qwen3.6. Worth noting how fresh this picture is: Google’s gemma4 family is live on Ollama and was last updated June 9, 2026. The local model shelf is moving monthly now.

hardware tiers, honestly

8GB RAM: the starter tier

You can run small models and they are better than people think.

• qwen3.5 0.8B (1.0GB) or qwen3.5 2B (2.7GB)

Good for summaries, drafts, quick questions, private journaling. Not good for long documents or serious reasoning. At this tier Odysseus works, but you will feel the ceiling. The other option is pointing it at a cloud API, which works but quietly gives away the whole privacy point.

16GB RAM: the real entry point

This is where local AI starts feeling like a daily tool, especially on Apple Silicon where the memory is unified.

• qwen3.5 4B (3.4GB) or qwen3.5 9B (6.6GB)
• gemma4 E2B (7.2GB) or gemma4 E4B (9.6GB), Google’s edge models, multimodal, 128K context

A 9B-class model in 2026 handles most of what normal people used ChatGPT for. This tier is most laptops sold in the last three years.

24GB RAM/VRAM: the sweet spot

• gemma4 12B, refreshed on Ollama this month
• qwen3.5 27B (17GB) starts to fit, with care
• gemma4 26B MoE (18GB), a mixture-of-experts model with only 3.8B active parameters, so it runs faster than its size suggests

This is where “local AI is a toy” stops being true.

32GB+: the no-excuses tier

• qwen3.6 27B (17GB) or gemma4 31B (20GB) run comfortably
• qwen3.6 35B (24GB) fits if you keep other apps closed

qwen3.6 specifically targets agentic coding and long thinking, the stuff people claim you need a cloud subscription for. At this tier you genuinely do not.

two footnotes that save you pain

Disk space is the silent requirement. Models are big files you keep. Two or three models in the 7-20GB range plus Docker images means you want 50GB+ free before this hobby starts, or your first error will be a full disk, not a slow model.

GPU VRAM beats RAM, but RAM still counts. A discrete GPU with 24GB VRAM runs these models fast. A Mac’s unified memory splits the difference and is why M-series laptops became the default local AI machine. A gaming PC with a 12GB card sits in between: great speed for models that fit, hard wall for ones that do not.

so why is everyone still confused?

Because nothing tells you this stuff up front. Odysseus ships a feature called Cookbook that scans your hardware and recommends models from a 270+ model catalog, which is exactly the right idea. But the project’s own ROADMAP admits the current ranking scores “everything almost the same” instead of prioritizing newer architectures and better hardware fit, and that failed model setups need to surface real error logs instead of mystery failures. The maintainers know, they wrote it down, and the fix is on the list. Week-two open source.

Meanwhile the failure mode for a normal person is brutal: you spend an evening on the install, Cookbook suggests a model, the model does not actually fit your memory, and your first experience of “owning your AI” is a frozen laptop. Most people do not try twice.

know before you install: the machine passport

This is the gap we are building HOMESTEAD around. Before HOMESTEAD installs anything, it builds a machine passport: a plain-language read of your actual hardware (memory, GPU, free disk) and a verdict on what your machine can run well, what it can run barely, and what it should not attempt. No VRAM math, no guessing, no frozen laptop.

Then the rest of HOMESTEAD is the part Odysseus proves people want: a one-click private local AI desktop app with private on-device memory. Nothing leaves the building.

HOMESTEAD is not out yet. The honest CTA is the waitlist at fuckbigtech.ai. If you read the tiers above and thought “I still do not know which one I am,” that is exactly the person we are building it for.

And if the tiers made sense and you want the full story of the local AI exodus, start with PewDiePie quit big tech AI. Here’s the easy way to follow him.