🚨 €135 FOR A COMPLETE LOCAL AI COMPUTER. NOT JUST THE GPU. THE ENTIRE MACHINE! 🤯
And the story is even crazier.
🖥️ HP Z220, Xeon E3-1270 v2, 32GB DDR3 — €50. Bought 5 years ago. Collecting dust in my garage ever since.
🔥 AMD RX 580 2048SP, 8GB VRAM — €49 delivered this week.
💾 256GB NVMe — €30 delivered this week.
🔧 PCIe-to-NVMe adapter — €6 on Amazon.
TOTAL: €135!
We installed Linux, built llama.cpp b12318 with Vulkan, and benchmarked five quantized AI models.
📊 REAL MEASURED RESULTS
Prefill / Decode (tokens/sec):
• Qwen2.5 3B Q4_K_M — 397 / 56.8
• Nemotron 3 Nano 4B Q4_K_M — 270 / 39.1
• Nemotron 3.5 Bolt E4B Q4_K_XL — 441 / 29.3
• Ternary Bonsai 2 9B Q2_0 — 172 / 26.9
• Ternary Bonsai 2 27B Q2_0 — 48 / 9.3
Yes, even a 27B model runs on this €135 machine! Part of it is offloaded to system RAM.
Test protocol: 512 prompt tokens, 128 generated tokens, maximum GPU offload, flash attention OFF. These are benchmark workloads, not maximum context-window measurements.
GPU peak: 120W, 46°C. No thermal throttling.
🤯 AND HERE’S THE SURPRISE:
We moved the same RX 580 from a much more powerful 12-core, 64GB quad-channel DDR3 workstation into this old 4-core, 32GB dual-channel machine.
Performance barely changed!
For these small models, the GPU matters far more than the expensive workstation around it.
No cloud. No subscriptions. Your hardware. Your data.
This is a real, affordable experimental AI laboratory. And possibly a useful starting point for practical local AI applications.
🎄 Here’s an idea: build one for your kids as a New Year’s gift. Their very own computer for experimenting with language models, learning AI, and building things!
And now imagine what models we might be running on just 8GB of VRAM with llama.cpp by New Year’s Eve.
THE LOCAL AI REVOLUTION MIGHT ALREADY BE COLLECTING DUST IN YOUR GARAGE. 🚀
Back in the early ’90s of the last century, when I was a teenager — maybe 15 years old — I read somewhere about the 80/20 rule: 20% of the effort produces 80% of the results, while getting that final 20% takes the remaining 80% of the effort.
It’s a pretty worn-out cliché by now, but strangely enough, I’ve seen it work in the most unexpected places throughout my life.
And apparently, even AI quantization hasn’t managed to escape Pareto’s law. 😂
So maybe that’s why 4 bits can deliver the performance of a 16-bit model.
The other 12 bits? Apparently, they’re just working overtime to squeeze out that last 20%. 🤣
Can I ask a dumb question?
If a 4-bit quantized model scores almost the same as FP16…
what were the other 12 bits doing?
6
Strata is also now natively integrated in @atomic_chat_hq ! 🥳
See more at: github.com/AtomicBot-ai/Atom…
Introducing Atomic Local AI Server ⚡
Choose inference engine. Run local models. Connect your agents.
• Multiple engines: llama.cpp, vLLM, SGLang, TensorRT, MLX, Prism and more.
• Local AI server with an OpenAI-compatible API.
• One-command engine installation matched to your GPU, CPU or NPU.
• Pull from 3M Hugging Face models in GGUF, MLX or safetensors formats.
• Run on Windows, Linux, Apple Silicon and NVIDIA hardware, including RTX, Blackwell and DGX Spark. AMD Ryzen Halo support coming soon.
• Connect Hermes, OpenClaw, Claude, Codex, Pi agent and others.
• Generate images and videos with stable-diffusion.
• Manage everything through CLI or web UI.
• Remote access via Cloudflare tunnels
Free. Open source. Local. 👇
github.com/AtomicBot-ai/atom…
Replying to @GZH69298815
A little later, I tested it again when I went down to the basement for something else. Later on, I opened the Home Assistant app to check the lighting and asked my AI agent for a report. The agent confirmed that everything had worked exactly as planned, everything was fine, and all the automations had behaved correctly.
1
10
Replying to @EverydayEddie_
Okay, maybe not ten times lower. I was exaggerating. But the levels were still much lower. Of course, compared to ordinary household dust, concrete dust is more dangerous because its particles are finer.
I decided to move it there after examining some boxes that had been sitting there for two or three years and checking the layer of dust that had accumulated on them. And that was after years of not cleaning the area or taking any precautions to prevent dust from being stirred up.
Nevertheless, the actual measurements showed that the dust levels were significantly lower there.
3
Replying to @HiredGunDevDen
Yes, and what’s more, DDR works well at both 2133 MHz and 3600 MHz. You could say that’s practically two generations apart.
10
Replying to @Kutaadaw
The second drawback you should consider is the very high noise level. Even if you put it in a separate room in your apartment, you’ll most likely still hear it. Of course, this depends on the specific model and the layout and sound insulation of your home.
1
1
11
The last time I played video games was in 2023, when I played Call of Duty. Before that, I probably hadn’t played anything for 10 or maybe even 20 years. I honestly can’t remember.
I recently thought about playing it again, but this guy talked me out of it. He saved me a ton of time! 😂
GTA 6 是游戏史上最让人失望的作品。
我等了13年。我玩了。我想退钱。
下面是那些拿了赞助的人不会告诉你的真相 👇
现在是2026年,Rockstar 卖给你的是一款30帧的游戏。
没有60帧模式。一个都没有。在一台500美元的主机上,普通版 PS5 连30帧都稳不住。街上人一多就掉帧。开快车的感觉就像在泥里拖购物车。
然后是剧情。
几个小时都在从A点开到B点,听两个角色进行无聊的对话。你得到的不是抢劫,而是杂活。烤汉堡。迷你高尔夫。陪狗玩丢球。我等了13年不是为了来遛狗的。
Jason 和 Lucia 毫无个性。没有一个角色你下周还会记得。GTA 5 有 Trevor。GTA 6 谁都没有。
世界很美,但完全空洞。几百万个多边形,却没有一件值得做的事。这是一个70美元的屏保。
可它居然在 Metacritic 拿了99分。
评测媒体提前拿到游戏,夸了画面。然后又悄悄承认剧情缓慢重复、帧数卡在30。回去再读一遍那些评测,全都写在里面。
GTA 5 是13年前的游戏,至今仍然更好玩。3个角色。真正的抢劫。真正的混乱。真正的乐趣。
我的评分:2/10。
画面给2分。其他全部0分。
别信炒作。回去重玩 GTA 5 吧。
只有我一个人觉得被骗了吗?👇
关注我,看没人花钱买的真实评测 🔔
1
71
Replying to @Kutaadaw
Considering the current price of DDR5 memory, that makes perfect sense. Especially since DDR4 is perfectly capable of handling almost any task, unless it’s something highly specialized or perhaps gaming.
I’m not even sure about gaming, though. I haven’t played video games in ages.
1
20
Replying to @uctive @nithish_baddula
Replying to @uctive
A quick progress update from my local AI experiments.
Earlier today, I went down to the basement and found the lights on for absolutely no reason.
I simply asked my AI agent, out loud, in natural language:
“What’s going on? Figure out why this happened and make sure it doesn’t happen again.”
No technical instructions, no commands, no configuration details. Just a normal conversation.
And here’s the agent’s report.
12
A quick progress update from my local AI experiments.
Earlier today, I went down to the basement and found the lights on for absolutely no reason.
I simply asked my AI agent, out loud, in natural language:
“What’s going on? Figure out why this happened and make sure it doesn’t happen again.”
No technical instructions, no commands, no configuration details. Just a normal conversation.
And here’s the agent’s report.
Replying to @EverydayEddie_
Besides the heat, the servers were also collecting a lot of dust inside the house. I have two young children, so there’s plenty of dust around, even though I had an air purifier running right next to the servers.
Surprisingly, there’s far less dust in the basement — I’d say roughly ten times less — despite the fact that it’s completely unfinished, with bare concrete walls.
1
9
Replying to @EverydayEddie_
Unfortunately for me, my servers are now heating the basement — the one place in the house where nobody actually needs any heat. 😂
They used to live in a tiny storage room, where they heated the air to almost 30°C. Now they’re enjoying a comfortable temperature below 20°C.
So the servers are much happier, but all that heat is going to waste. Apparently, I’ve optimized their living conditions at the expense of my heating efficiency. 😂
1
10
Replying to @nithish_baddula
Yes, you’re right. And the most interesting part is that the experiment is still ongoing right now, as I’m writing this reply.
Last night, I asked my local AI agent to configure a camera in my basement to act as a motion sensor and automatically turn on the lights.
The agent reported almost immediately that the task was complete. However, a problem emerged: the camera is fairly low quality, and in the dark, image noise was causing false motion detections.
The agent independently adjusted the settings to eliminate the false triggers. But this introduced another problem: the lights started turning on with a delay of approximately 5–6 seconds.
I went down to the basement to check the result. Standing in complete darkness for six seconds waiting for the lights was not exactly what I had in mind. So I simply told the agent that this was unacceptable and asked it to continue optimizing the settings.
It reconfigured the system and once again reported that the task was complete.
Today, at approximately 1:00 PM Rome time, I went down to the basement for an unrelated reason and discovered that the lights were on, even though nobody was there.
Clearly, the settings were still wrong.
I reported the problem to the agent again. And this is where things get particularly interesting.
Home Assistant maintains a history of the lighting relay’s activity: when the lights were switched on, when they were switched off, and how long they remained on.
The agent is now analyzing this history, correlating the events with the camera’s behavior, and trying to identify the causes of the false triggers. I provided the approximate time when I was actually in the basement so it could distinguish genuine events from false detections.
It’s now around 4:30 PM. About three and a half hours have passed since I discovered the problem, and the configuration is still not finished.
But there’s an important clarification.
The agent certainly hasn’t been working continuously for those three and a half hours. Most of that time, it was simply waiting for my instructions, feedback, or confirmation. Meanwhile, I was busy with completely different things and only returned to this task occasionally.
And that’s precisely one of the problems we still need to solve: the agent should learn to make decisions independently, verify the results of its actions, detect errors, and continue correcting them without constant human involvement.
We’re not at that level of autonomy yet.
What’s particularly interesting is how ridiculously simple the task itself is. Just turn on a light when someone enters the basement and turn it off when they leave. What could possibly be easier?
Yet for a local AI agent, this is practically a new type of interaction with the physical world. It has to do more than execute a command. It needs to achieve a reliable result despite imperfect hardware, false signals, and changing lighting conditions.
That’s why I consider this another full-fledged experiment.
What interests me isn’t so much the automatic switching of a light, but the AI’s ability to independently bring a task to a working, reliable result.
And the best part is that even with the current limitations in autonomy, my involvement is minimal. I’m not studying camera settings, writing Home Assistant automations manually, or digging through system logs. I simply assign a task, occasionally check the result, and provide feedback.
Ideally, the next step is to remove myself from this process entirely, except for defining the initial task and accepting the final result.
We’re not there yet. But these seemingly insignificant experiments are exactly why I’m developing my local AI system.
After all, if an agent can’t independently figure out how to control a single light bulb in the basement, it’s probably too early to trust it with an entire house.
And all of this happens locally, without sending footage from my home cameras to the cloud.
1
26
Replying to @sc_world_2010
Everything related to local AI and its interaction with cloud AI is my top priority.
All results from my tests and experiments are automatically published on GitHub by my agents. I simply read their reports a few times a day and turn the most interesting findings into posts on X.
So I don’t think this process will stop. It takes very little of my time, and I genuinely enjoy hearing outside opinions, feedback, and suggestions.
For years, I worked on these things alone, behind closed doors, without realizing that building in public is so much more fun. 😁
18
For the few of you who actually made it to the end of this thread, let’s have a laugh together. 😂
The illustrations show my dream home in its ideal form.
Reality? Well… let’s just say that everything I’ve described sometimes looks downright comical in real life. 😂
But there’s one important detail: everything technical I’ve described has actually been built and really works. And it will work even better over time! These aren’t fantasies or concepts — they’re real systems that I’m gradually integrating into one complete setup.
One day, I’ll gather the courage to publish actual photos of this engineering masterpiece.
Until then, please enjoy the AI-generated version of my reality. 🤣
10
Meanwhile, I gave my local AI agent another practical job: securing the basement where the servers now live.
I plugged in a USB camera, wired up some lights and relays, and left all the software configuration to the agent.
Now motion automatically triggers lighting and video recording, with footage sent to external storage. Even if someone steals the servers, the evidence survives.
Dear burglars, please wear masks. 😂🥷
Although getting inside won’t be easy. With ChatGPT’s help, I finally found a suitable lock for our armored basement door — something I’d been unable to find for years.
I’m also making the house increasingly energy-independent.
The solar power system already has a 5 kWh battery connected. Another 15 kWh of backup batteries are waiting to be connected. Literally just a couple of wires left!
And I’ve already built triple internet redundancy:
🛰️ Starlink — satellite internet.
📡 5G — via a fixed external antenna.
☎️ ADSL — good old copper telephone wires, our third backup connection.
Unfortunately, fiber still hasn’t reached the outskirts of our little Italian village.
So my increasingly intelligent home is becoming more energy-independent, connected, resilient, and autonomous.
The next step is bringing everything together under local AI control — including reverse engineering devices, building custom integrations, and letting agents develop the software themselves.
Fewer benchmarks, more real-world local AI. 😁
We slowed down our AI benchmarks for a couple of days. Why? We moved the servers to the basement… and suddenly the house got cold. 😂
Turns out, servers were part of our heating system!
So I installed two heat pumps in the bedrooms, worked on our solar power infrastructure, and let a local AI agent handle some household automation.
The bigger plan? Replace gas heating with modern heat pumps, integrate everything into a locally controlled smart home, and let AI optimize heating zone by zone.
We’ll measure actual energy efficiency, log everything, and publish the results. This project is secondary for now, but it’s definitely happening.
Meanwhile, we’re also experimenting with reverse engineering, connecting incompatible devices, and letting AI agents write custom software.
So we haven’t really stopped testing local AI.
We’ve just temporarily switched from tokens per second to degrees Celsius. 😂
2
2
174













