Replying to @murasametech
Just blow it out once every couple of weeks. It only takes a minute. Just don’t hold a thing like that too close 🤣
5
14
8,195
🇪🇺 FOR SALE — European Union | Milan, Italy
I’m selling two powerful Apple computers, both purchased in spring 2026, in excellent condition and available for immediate delivery.
🖥 Mac Studio M3 Ultra
• 28-core CPU / 60-core GPU
• 256GB unified memory / 2TB SSD
• Price: €9,900
• AppleCare+ warranty and accidental damage protection until May 19, 2027, with an option for annual renewal.
💻 MacBook Pro 14” M5 Max
• 18-core CPU / 40-core GPU
• 128GB unified memory / 2TB SSD
• Price: €6,900
• AppleCare+ warranty and accidental damage protection until July 2, 2029.
✅ Both available immediately in Milan, Italy.
✅ Prices include VAT.
✅ Company invoice with VAT separately itemized.
✅ Payment by bank transfer to the company account.
✅ Prices negotiable.
Why am I selling?
My AI startup is growing, and after receiving cloud computing credits through startup programs from Amazon AWS, Google Cloud and OVHcloud, I now have more computing capacity than I currently need.
I’m selling this surplus hardware to free up funds for continuing my research, development and growth of the startup.
🙏 A HUGE FAVOR TO EVERYONE: PLEASE HELP ME SELL THESE COMPUTERS!
I really need the funds to keep working on my project and developing my startup.
Even if you’re not interested in buying, please repost and share this with anyone who might be interested.
Your support would mean a lot to me. Thank you! ❤️
🚨 €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. 🚀
By the way, complete PC systems with an RTX 5060 Ti 16GB cost around $1,500 in the US. For just over $2,000, you can already get a system with an RTX 5070 Ti 16GB.
And Micro Center’s PowerSpec offers extended warranty coverage of up to 3 years for just $220 extra.
These systems also use standard, replaceable components and have access to an established service network.
Of course, WiCi One offers additional wireless GPU technology, but the difference in price, repairability, and long-term support is hard to ignore.
🔗 microcenter.com/product/7142…
1
89
WiCi One: Why Does a Device Costing $12,199 Come with Only a One-Year Warranty?
Let’s compare the standard manufacturers’ warranties for the components of a high-quality PC:
🎮 NVIDIA RTX 5090 GPU (32 GB) — 3 years.
🖥️ ASUS, MSI, Gigabyte motherboards — typically 3 years, sometimes up to 5 years.
🧠 Kingston, Corsair, G.Skill RAM — limited lifetime warranty.
⚡ Seasonic, Corsair, be quiet! power supplies — up to 10–12 years.
💾 Samsung, WD, Kingston SSDs — typically 3–5 years.
🔲 Intel or AMD processors (boxed versions) — typically 3 years.
🖥️ High-quality PC cases — typically 2–3 years.
🚨 Now consider the WiCi One with an RTX 5090, priced at $12,199 — just a one-year warranty for the entire device.
At the same time, the manufacturer does not disclose the system RAM capacity, motherboard model or form factor, whether components can be replaced independently, or any long-term commitments to supporting its proprietary GPU driver.
❓ This raises an obvious question: What’s wrong with this device?
Why does a computer built from components whose manufacturers typically offer multi-year warranties come with only one year of coverage?
And what happens in two or three years if a non-standard motherboard fails or the startup stops supporting its software?
I’m not claiming that the device is unreliable. The technology is interesting, and the engineering concept deserves attention.
But at this price, I’d expect more than innovation alone. I’d also expect transparent specifications, repairability, and guarantees of long-term usability
🔗 Official website: wici.ai/shop/wici-one
2
3
561
Replying to @OVeselenko @since1795
Посмотрел их сайт:
1. Дешевый миниписи с неизвестным объемом RAM даже при предзаказе
2. Продаж реальных нет
3. Гарантийное обслуживание - не проверено, риск крайне высокий, так как сборка кастомная
4. Дизайн абсолютно другой
5. Версии с 5090 в предзаказах нет
6. Операционная система неизвестна
7. Цена высокая на такие комплектующие
8. Применение слишком узкое
9. Легко заменяется нормальным системником где гарантия не один год а два или три и запчастей навалом
Шансы провала проекта и гарантийного обслуживания - 99%
вычеркиваем 🤣
2
1
34
🔥 What matters more for local AI: DDR5 RAM or GPU VRAM?
We tested how increasing system RAM from 32GB to 64GB and available VRAM from 16GB to 24GB affects a model that doesn’t fit entirely into either memory pool.
Our machine: P2005, RTX 5090 Laptop 24GB, 64GB system RAM.
Model: Qwen3.8-Flash-Next, IQ3_S quantization, 82 GiB (Qwen3.8-FN-IQ3_S.basalt).
Engine: Basalt v0.2.1, released November 4 — just 9 days ago!
🧪 Round 1 — More RAM, VRAM locked to 16 GiB:
32GB → 83 tok/s (peak 87)
48GB → 97 tok/s (peak 104)
64GB → 105 tok/s (peak 109)
Twice the RAM. About 27% more speed.
Useful, but we wanted more! 😁
🚀 Round 2 — More VRAM, with 64GB RAM:
16 GiB → 105 tok/s (peak 109)
20 GiB → 144 tok/s (peak 152)
~24 GiB → 187 tok/s (peak 189)
💥 Almost 80% faster!
Expert-cache hits jumped from 75% to 94%. Less SSD traffic, more speed.
And we hadn’t changed a single hardware component!
We thought that was a pretty good result.
Then we checked our inference engine… 🤦
Our 9-day-old Basalt v0.2.1 was already 469 commits behind!
So we upgraded to Basalt v0.3.3, released November 12 — the latest release today.
And the experiments continued!
⚡ Best short-prompt decode: 189 → 204 tok/s
⚡ Cold-start decode: 112 → 127 tok/s (+13%)
⚡ 27.5K-token prefill: 3,263 → 3,768 tok/s (+15%)
⚡ Long-context decode: 137 → 165 tok/s (+20%)
The 204 tok/s result was a peak in a tuned test. Typical warm decode remained around 185–187 tok/s, so the upgrade’s real advantage is long-context performance, not sustained short-prompt speed.
For perspective, the older engine even logged a 353 tok/s peak in real production traffic — a different workload, not a directly comparable benchmark.
That’s what makes local AI benchmarking so interesting: memory limits, cache behavior, workload and engine versions can completely change the numbers.
Same computer. Same 82 GiB model. Lots of experiments. And we’re still learning! 🚀
A full year of working full-time with artificial intelligence, AI agents, a huge zoo of models, and all kinds of hardware has led me to a rather unexpected workspace setup.
I’d like to share my experience. Maybe someone will find it interesting. 😄
I have around 10 computers currently undergoing testing as prototypes for local AI systems, ranging from cheap old machines to some pretty powerful configurations. I’m constantly experimenting with models, inference engines, agents, and different hardware setups.
But the most interesting part is how my own working habits have changed.
I have three physical workspaces. And a fourth one, which is actually my main workspace.
1. Kitchen / Living Room
An old laptop from around 2018, with 16 GB RAM, an old Intel CPU, and some GPU I can’t even remember.
I once tried to sell this laptop for €200, but nobody wanted it. Now it’s my main computer whenever I actually need a screen.
This workspace exists because I have two young children: my son is 6 and my daughter is 9. I need to keep an eye on what they’re doing.
2. Bedroom
An iPad with a keyboard. I rarely use it, mostly when I need privacy, peace, and concentration.
The same iPad also serves as my mobile terminal when I leave home. I take it with me for work, presentations, and demonstrations of my AI systems.
3. Basement
That’s where my experimental local AI laboratory and prototypes are located.
It’s cooler down there, with less dust. There’s a regular desktop computer with a monitor, also running a local AI model. That model mainly serves as an additional brain for handling tasks.
I rarely sit at the computer itself. I mostly go downstairs for physical work: swapping GPUs, fixing hardware, checking connections, restarting equipment, or testing something on-site.
4. My Armchair
My real, main workspace. 😂
A smartphone in my hand. No keyboard, no mouse, no big monitor.
And this is where I actually do most of my work.
Why did I end up with such a strange configuration? That’s what the next post is about.












