My assistant's sensory cortex is a games console
A Steam Machine with 8 GB of VRAM, running sight, hearing, voice, imagination and memory for a second brain. One model at a time, no cloud, no API bills.
The most wanted gaming console of 2026 lives in my living room. By day it plays games. By night it is my AI assistant's sensory cortex.
Valve's new Steam Machine. 8 GB of VRAM, which is nothing by AI standards. Here is what Morpheus, my second brain, runs on that one small box:
- Sight. Qwen 2.5 VL 7B reads and OCRs any image I show it.
- Imagination. SDXL draws its illustrations, two checkpoints, ESRGAN upscaling to 2K.
- Voice. ACE-Step 3.5B sings actual songs, with lyrics it wrote itself.
- Hearing. Whisper transcribes any audio, on the CPU, for free.
- Thought. Qwen 2.5 7B chats and Qwen 2.5 Coder 7B writes code, fully offline.
- Memory. nomic-embed-text for semantic search, plus a cited RAG working through 500 episodes of a business podcast. It quotes them with timestamps.
The trick is a small custom MCP server on the box. One model fits at a time, so it boots the right one on demand, batches calls by mode, and enforces the safety caps that keep the machine alive. The 30 second music cap is not a vibe, it is measured: about 165 MB of VRAM per second of audio. Every capability is a tool over MCP, so any agent in my setup can use the box as a sense.
No cloud. No API bills.
So here is my question. Should I open source steam-machine-ai-stack? The server, the model manager, the safety rails, the lot. If you would run one on your own box, tell me in the comments.
Want to hear what Morpheus sounds like? Check out koik.ai
My own first comment
Before anyone asks what it takes to run this: nothing in the stack is actually Steam Machine specific. It is ComfyUI plus Ollama plus one small custom MCP server that boots and unloads the models, and it would run on any Linux box with 8 GB of VRAM. The console is just the best looking server I have ever owned. If the interest is real, I will write up the AMD ROCm quirks too.
Where this first appeared
Written for LinkedIn and posted there on . Read it on LinkedIn.