Vik | Practical AI & Tech (@uctive) header
Vik | Practical AI & Tech (@uctive) profile picture

Private local AI server for small businesses: your data stays inside your company. https://nitter.gq/t.co/dnhFJ2Yoqh

Como, Italy
Joined December 2011
🚨 €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. 🚀
5
6
1
22
3,492
Vik | Practical AI & Tech retweeted
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…
6
7
1
123
3,900
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
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
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. 😁
2
1
103
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.
1
1
1
54
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
1/5. Why am I switching from DDR5 to DDR4 at the end of 2026? And it’s not about money. @coldniko Strata changed my understanding of what hardware is actually needed to run local AI. So much so that I’m now moving my entire home and business infrastructure from a modern DDR5 server with ultra-fast PCIe 5.0 NVMe drives to a much older DDR4 platform. I already own both DDR5 and DDR4 hardware. I’m not trying to save money. My current server works beautifully, and I’m completely satisfied with it. So why am I voluntarily downgrading? Here’s the story.
7
6
1
90
12,554
5/5. I want to use what I’m going to sell Today, I’m migrating all my main services, databases, Home Assistant, Docker containers, RAG, accounting, and business workflows to DDR4. I’ve published the hardware configuration before, but I’ll share it again separately. The only significant difference from the entry-level customer system is my relatively expensive SAS controller and two 20 TB HDDs in RAID 1 for data storage. My four-drive PCIe 5.0 RAID 0 array, delivering 30–35 GB/s, will be gone. And I’m genuinely curious whether I’ll even notice the difference in everyday work. I don’t want to sell someone a €1,000 server while continuing to use an expensive machine myself because the cheaper one supposedly isn’t good enough. I want to live and work on the same class of hardware I’m offering my customers. If something breaks, I’ll be the first to experience it. If everything works beautifully, that’s far more convincing evidence than any synthetic benchmark. So at the end of 2026, I’m voluntarily moving from DDR5 to DDR4. Not to save money. To test my business idea in real life.
1
8
1,228
BONUS 6/5 — My actual hardware downgrade: DDR5 → DDR4 Here’s the configuration I’m moving my entire home and business infrastructure to. BEFORE — CLARYEL01: * CPU: Intel Core Ultra 7 270K Plus, 24 cores * Motherboard: ASUS Pro Q870M-C * RAM: 64 GB DDR5 * Storage: 4× Lexar NM1090 PRO 2 TB, PCIe 5.0 * RAID 0: 30–35 GB/s read throughput * Additional SSD: Samsung 990 PRO 2 TB * GPU: none AFTER — P1005 (DDR4): * CPU: Intel Xeon E5-2690 v4, 14 cores / 28 threads * Motherboard: ASUS Z10PA-U8 * RAM: 64 GB DDR4 ECC * GPU: NVIDIA RTX 3050 8 GB * SSD: Patriot Viper VP4300L 1 TB NVMe * SAS controller: LSI SAS3008 12G * Storage: 2×20 TB SAS HDD, RAID 1 * OS: Ubuntu 24.04 LTS WHAT’S MOVING? Around 40 Docker containers and supporting services: Home Assistant, Zigbee, Mosquitto MQTT, ERPNext, Odoo, OpenProject, PostgreSQL, MariaDB, MongoDB, Qdrant, Meilisearch, SeaweedFS, Valkey, Redis, LibreChat, LiteLLM, Activepieces, Ollama, ClamAV, agent infrastructure, knowledge APIs, integrations, monitoring, document storage, backups, and business automation. Plus Hermes Agent, with 40 scheduled jobs. The primary AI inference engine remains on a separate machine. I’m moving from a modern 24-core DDR5 platform with extremely fast PCIe 5.0 storage to an older 14-core Xeon with DDR4 and a modest 8 GB GPU. This isn’t about saving money. I already own the hardware. I want to find out whether a €1,000-class local AI platform can reliably handle my actual home and business workloads. And I’m going to find out by using it myself, every single day.
4
2
11
1,220
I completely agree — give it a try! I’d just add that Q4 is great, but Q3 is already good enough for many practical tasks. For about three weeks now, maybe even longer, I’ve been using just one model locally: Qwen 3.8 Next IQ3_S. And I’ve been running it on all kinds of hardware, from €500 systems to €10,000 machines.
dear 8gb, 12gb, 16gb and 24gb vram owners, i know people in the local ai space might discourage you, saying don't use q4 and q4 kv, it trades away quality and sucks and all that, but hear me out loud, i'm going to be very clear the ones saying that own datacenter grade chips and can afford full precision, and as a small gpu owner you don't have a choice i'm out here in the field running everything from 8gb cards to 256gb of unified memory, and i know what i'm saying don't get sold, try it and find out, it's not that bad at all
2
16
2,879
He’s coming back. This is a huge sign for us. And I don’t think the board of directors — those nobodies nobody has ever heard of — will dare stand in his way.
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
1
192
Vik | Practical AI & Tech retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
58,685
119,496
24,165
1,142,200
54,064,663
Vik | Practical AI & Tech retweeted
It started as a weekend project to learn local inference. Now, it's a growing community. Thank you for 1,000 GitHub stars and 100+ forks! A special thanks to all the contributors and everyone who spent time testing and sharing ideas. This is just the beginning!
3
4
1
16
784
Don’t worry, I’ll probably publish a very negative post about local AI soon, once I’ve accumulated enough frustrations to write about. But that doesn’t mean local AI is unnecessary. Quite the opposite. It proves why we need it, while also reminding us that it’s a tool with both advantages and disadvantages. And there are an enormous number of advantages. As for people spending too much money on hardware, how is that different from any other hobby? Some people spend fortunes on cosmetics and clothes. Others collect watches, cars, or all sorts of things. Some enthusiasts pour incredible amounts of money into smart homes, solar panels, computers, and countless other projects. Everyone has their own obsession. What seems like an absurd expense to one person can be a source of genuine happiness to another. And if someone enjoys the process, I think that’s perfectly fine, even if they never achieve any practical result. As for people not sharing their negative experiences, that’s perfectly understandable too. Nobody wants to look foolish in front of others, and many are understandably afraid of criticism, ridicule, or online hate. I think all of this is completely normal. It’s simply human nature, and we see the same behavior in practically every field. Yet despite all these imperfections, people keep experimenting, learning, making mistakes, and sharing discoveries. And together, humanity keeps moving forward. The progress we’re witnessing is absolutely incredible.
Replying to @uctive
I would be there are many times more people they bought hardware and never did anything worthwhile. X is full of people going crazy but no one posts negative stories.
1
1
493
RTX 5090 Laptop 24GB + Basalt v0.4.1 — express benchmark: successes and failures. At our followers’ request, we tested the dense Qwen3.8-27B-QUASAR-NVFP4 model with contexts up to 256K tokens. 🚀 Up to 121 tok/s with DFlash 2. But long contexts revealed some interesting limitations. 🧵
4
1
7
1,019
6/7 🔬 ADDITIONAL FINAL RESULTS — RTX 5090 Laptop 24GB + Basalt v0.4.1 Our overnight benchmark is now complete! This is the first of two additional posts following our original five-part thread. The full report, raw logs, configurations, benchmark scripts, and four charts are now available on GitHub. We discovered several interesting things and, most importantly, established the actual context limits within 24GB of VRAM. ━━━━━━━━━━━━━━━━ 🎯 1. THE REAL VRAM CONTEXT LIMITS Our previous estimates — 221K tokens without DFlash and 173K with DFlash 2 — turned out to be too optimistic. They were based only on KV-cache memory calculations and did not fully account for the engine’s actual memory allocation requirements. We have now established the real limits through binary search, repeatedly starting the engine with different context sizes and recording successful starts and allocation failures. ✅ WITHOUT DFLASH (–spec 0) • Maximum working context: 203,776 positions • First rejected configuration: 205,824 positions ✅ WITH DFLASH 2 (–spec 16) • Maximum working context: 140,288 positions • First rejected configuration: 142,336 positions These are verified maximum engine context allocations with the entire KV cache residing in VRAM. DFlash 2 requires approximately 2 GiB of additional GPU memory for its speculative drafting mechanism. This significantly reduces the available context capacity. ━━━━━━━━━━━━━━━━ 🚀 2. PERFORMANCE AT THE ACTUAL VRAM LIMITS 📦 PLAIN DECODING — NEAR 204K • Maximum configured context: 203,776 positions • Actual input prompt: 197,886 tokens • Prefill: 2,476 tok/s • TTFT: 80.04 seconds • Generation: 23.5 tok/s • Peak VRAM: 23,160 MiB Speed across four consecutive quarters of one continuous response: 23.9 → 23.5 → 23.5 → 23.5 tok/s Even with almost 198K input tokens, generation speed remained remarkably stable throughout the response. ⚡ DFLASH 2 — NEAR 140K • Maximum configured context: 140,288 positions • Actual input prompt: 134,561 tokens • Prefill: 3,136 tok/s • TTFT: 43.09 seconds • Generation: 64.7 tok/s • Draft acceptance: 63% • Peak VRAM: 23,498 MiB Speed across four quarters: 55.0 → 78.5 → 76.9 → 63.6 tok/s In both tests, the entire KV cache remained in GPU memory. ━━━━━━━━━━━━━━━━ 🏆 3. A NEW SPEED RECORD We previously reported a maximum of 121 tok/s in a standard chat task with a 4K context. However, testing different task types revealed substantially higher throughput. ⚡ JSON GENERATION, 4K: • Whole-response generation speed: 194.3 tok/s • Fastest individual quarter: 264.3 tok/s • Draft acceptance: 69% Additional DFlash 2 results: 📋 JSON, 32K — 144.5 tok/s 💻 Code analysis, 32K — 106.3 tok/s 💻 Code analysis, 64K — 73.4 tok/s These are different tasks with different prompt content, so their results should not be treated as directly interchangeable benchmark scores. However, they clearly illustrate how strongly DFlash 2 performance depends on the workload and draft acceptance rate. We also tested --suffix-draft 3: • 4K: 119.9 vs 121.0 tok/s • 66K: 91.3 vs 94.1 tok/s No performance improvement was observed beyond normal measurement variation. ━━━━━━━━━━━━━━━━ ➡️ NEXT — 7/7: What happens beyond VRAM capacity, long-context retrieval accuracy at 256K, final conclusions, and the complete GitHub repository.
1
62
7/7 🏁 FINAL POST — BASALT v0.4.1 BENCHMARK CONCLUSIONS This is the second and final additional post in our thread about testing the dense Qwen3.8-27B-QUASAR-NVFP4 model on an RTX 5090 Laptop with 24GB VRAM. ━━━━━━━━━━━━━━━━ 🐢 4. WHAT HAPPENS BEYOND VRAM CAPACITY? At 256K, the entire KV cache no longer fits in GPU memory. We used --kv-resident to move part of the KV cache into system RAM. Results: 📦 Plain decoding — 4.7 tok/s ⚡ DFlash 2 — 5.1 tok/s DFlash 2 acceleration dropped from 3.2× at 131K to just 1.09× at 256K. The reason is the bandwidth limitation between system RAM and the GPU. Our laptop supports PCIe 5.0 x16, but the actual negotiated connection runs at PCIe 4.0 x8 — 16 GT/s per lane. Measured CPU→GPU bandwidth: 14.4 GB/s. Once KV-cache transfers become the main bottleneck, speculative decoding offers almost no advantage. This does not mean the model is incapable of processing 256K tokens. It means the performance of this particular configuration is primarily limited by data transfer. ━━━━━━━━━━━━━━━━ 🧠 5. LONG-CONTEXT RETRIEVAL ACCURACY We embedded five control records into long documents and attempted to retrieve them. The initial instruction was: “Find and list all control records.” Result: 0/5. However, we discovered a methodological issue in the follow-up request, so this result cannot be considered a reliable measurement of the model’s ability to independently identify every record. We then performed targeted retrieval tests, asking for each control record by its exact identifier. Results: ✅ 131K context — 5/5 correct. ✅ 256K context — 5/5 correct. Even at 256K, with part of the KV cache residing in system RAM, the model correctly retrieved all five requested values. This distinction is important. Technically accepting a long context, retrieving a specific fact by a known identifier, and independently identifying all relevant facts are different capabilities. We confirmed successful targeted retrieval, but we have not established reliable free-form discovery of every relevant fact. ━━━━━━━━━━━━━━━━ 📊 6. FINAL BENCHMARK RESULTS On the RTX 5090 Laptop 24GB, running Basalt v0.4.1 with the dense Qwen3.8-27B-QUASAR-NVFP4 model: 🏆 Maximum generation speed on an individual task: 194.3 tok/s ⚡ Maximum speed during an individual response quarter: 264.3 tok/s 🚀 Standard chat: up to 121 tok/s ✅ Maximum fully GPU-resident context: • Without DFlash — 203,776 positions • With DFlash 2 — 140,288 positions 📦 With an actual input prompt of 197,886 tokens, plain decoding maintained 23.5 tok/s. 🐢 At 256K, with partial KV-cache offloading to RAM, generation speed dropped to approximately 5 tok/s. Plain decoding remained remarkably stable throughout each continuous response. DFlash 2 behaved differently: generation speed could vary significantly within a single response, sometimes increasing substantially after the first quarter. ━━━━━━━━━━━━━━━━ 💡 7. THE MAIN TAKEAWAY The theoretical memory footprint of a KV cache is not the same as the maximum context an inference engine can actually allocate in VRAM. And even a powerful GPU cannot guarantee high performance when data must constantly move across a relatively slow interconnect. This is why benchmarking the entire system matters more than looking at isolated GPU specifications. Real engine. Real model. Real hardware. Real workloads. ━━━━━━━━━━━━━━━━ 📂 FULL REPORT AND ALL SOURCE DATA — GITHUB github.com/ViktorIXAI/basalt… Basalt v0.4.1 Commit: d3b7574 Model: Qwen3.8-27B-QUASAR-NVFP4 🔬 Benchmark completed. Our main local AI system on P2005 has been restored and is operating normally. Thanks to everyone who suggested these tests and asked technical questions. More experiments ahead.
1
83
My first piece of advice to anyone getting started with local AI: DON’T BUY HARDWARE. First, rent the GPU you’re considering buying from a cloud provider. A few hours will cost you just a few euros. Get a basic ChatGPT or Grok subscription, for example, and ask it to help you configure that rented GPU, deploy your model, and run it in the cloud. Test everything yourself. You might spend a few dozen euros, but you’ll have a much better idea of whether you actually need local AI, what inference speeds to expect, and whether those speeds are acceptable for your tasks. There’s an even easier option. Install Hermes and connect AI models through their APIs. Test their capabilities, reasoning quality, and response speeds. Then compare those results with published hardware benchmarks to estimate what performance you could realistically achieve on different local systems. Only after doing all of this should you even consider buying hardware. Don’t gamble on expensive equipment. Otherwise, you risk wasting not only money but also time. And time is the most valuable resource any of us has.
if i could give 1 piece of advice to anyone starting local ai for personal inference, learn four tools right after you buy the gpu > llama.cpp if it's one gpu > vllm if it's 2+ gpus > tmux so the serve outlives ssh > tailscale to reach it anywhere launch every model inside tmux, close the terminal or drop ssh and it keeps running, and if you can, don't hesitate on a tailnet either, your own api on your phone and laptop from anywhere with no ports open to the internet learn these and the road opens, you level up as you go, no rush, just go and get em
22
24
3
318
26,944