{"id":29345,"date":"2026-10-23T22:24:36","date_gmt":"2026-10-23T22:24:36","guid":{"rendered":"https:\/\/www.websiteradar.net\/?p=29345"},"modified":"2026-10-23T22:24:36","modified_gmt":"2026-10-23T22:24:36","slug":"local-ai-hardware","status":"publish","type":"post","link":"https:\/\/www.websiteradar.net\/en\/local-ai-hardware\/","title":{"rendered":"Local AI hardware in 2026: how to choose a computer for running LLMs at home"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">Want to run language models on your own machine \u2013 no subscription, no data leaving your desk, and no internet connection required? Then the CPU and the &#8220;TOPS&#8221; sticker on the box matter far less than two other things: <strong>how much memory the model can use<\/strong> and <strong>how fast the computer can read that memory<\/strong>. This guide explains how local AI actually runs on a computer, how much memory popular models need, and how today&#8217;s hardware performs in practice \u2013 from a graphics card you add to the PC you already have, to a workstation for a whole team. Every device comes with verified numbers, sources and a clear verdict on who it suits; some sections also have a short video.<\/p>\n\n\n\n<div class=\"wp-block-group dw-summary has-base-2-background-color has-background\" style=\"border-radius:16px;padding-top:24px;padding-right:28px;padding-bottom:24px;padding-left:28px\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-container-core-group-is-layout-bba1aa89 wp-block-group-is-layout-constrained\">\n\n<h3 class=\"wp-block-heading\">The 30-second version<\/h3>\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory decides everything.<\/strong> The whole model has to fit into fast memory that the GPU can reach. At 4 bits, budget roughly <strong>0.6&nbsp;GB per billion parameters<\/strong>.<\/li>\n\n\n<li><strong>Memory bandwidth sets how fast the model writes<\/strong>, compute sets how fast it reads a long prompt. That is why a Mac writes faster than a DGX Spark, while the Spark reads long documents faster.<\/li>\n\n\n<li><strong>Best value right now:<\/strong> Mac Studio M5 Max with 128&nbsp;GB (fast output), Ryzen AI Max+ 395 with 128&nbsp;GB (cheapest large memory), RTX 5060 Ti 16&nbsp;GB (cheapest sensible entry point).<\/li>\n\n\n<li><strong>An &#8220;AI PC&#8221; with an NPU is not a local AI computer.<\/strong> Neither Ollama nor LM Studio uses the NPU.<\/li>\n\n<\/ul>\n\n<\/div><\/div>\n\n\n\n<p class=\"dw-disclosure has-small-font-size wp-block-paragraph\">We did not test these devices ourselves \u2013 the numbers come from the sources listed at the end, and community measurements are marked as such. Prices are US list prices with their date; memory shortages pushed many of them up during 2026.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"podle-rozpoctu\">What to buy: the quick picks<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">If you don&#8217;t want to read the whole guide, here is the shortcut. Every recommendation is covered in detail below, some with a video.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1963\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-scaled.webp\" alt=\"What to buy for local AI, from entry level to workstation\" class=\"wp-image-29337\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-300x230.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-1024x785.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-768x589.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-1536x1178.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-2048x1571.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.en_-610x468.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Quick picks by class, from the cheapest sensible entry point to a team workstation.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Class<\/th><th>Pick<\/th><th>Models it handles<\/th><\/tr><\/thead><tbody><tr><td>Entry<\/td><td><a href=\"#rtx-5060-ti-16gb\">RTX 5060 Ti 16&nbsp;GB<\/a> in your current PC<\/td><td>8\u201320B (Qwen3 14B, gpt-oss-20b)<\/td><\/tr><tr><td>Starter<\/td><td><a href=\"#mac-mini-m6\">Mac mini M6 with 32&nbsp;GB<\/a> or <a href=\"#rtx-5070-ti-vs-5080\">RTX 5070 Ti<\/a><\/td><td>up to 14B at speed, Qwen3 30B-A3B<\/td><\/tr><tr><td>64&nbsp;GB<\/td><td><a href=\"#mac-mini-m5-pro\">Mac mini M5 Pro with 64&nbsp;GB<\/a><\/td><td>20\u201335B, a quiet home server<\/td><\/tr><tr><td>128&nbsp;GB, best value<\/td><td><a href=\"#ryzen-ai-max-395\">Ryzen AI Max+ 395 with 128&nbsp;GB<\/a><\/td><td>large MoE models up to 120B<\/td><\/tr><tr><td>128&nbsp;GB, fastest output<\/td><td><a href=\"#mac-studio-m5-max\">Mac Studio M5 Max with 128&nbsp;GB<\/a><\/td><td>gpt-oss-120b at over 100 tokens\/s<\/td><\/tr><tr><td>CUDA developer box<\/td><td><a href=\"#dgx-spark\">DGX Spark or another GB10 machine<\/a><\/td><td>CUDA development, coding agents<\/td><\/tr><tr><td>Maximum memory<\/td><td><a href=\"#mac-studio-m5-ultra\">Mac Studio M5 Ultra with 256&nbsp;GB<\/a><\/td><td>100\u2013250B models at home<\/td><\/tr><tr><td>Team workstation<\/td><td><a href=\"#rtx-pro-6000\">RTX PRO 6000<\/a><\/td><td>70\u2013120B for a whole team<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"proc-lokalni-ai\">Why run AI locally \u2013 and what not to expect<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A local model runs on your own computer. Your data never leaves the desk, you pay no monthly subscription or per-token fees, the model works offline, and nobody can change or retire it overnight. That makes it a good fit for sensitive documents, coding assistants working on company code, automations that send thousands of requests a day \u2013 or simply for the fun of experimenting.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">There are two catches. First, the open models that fit on home hardware are smart, but not at the level of the best cloud services. Second, the hardware is not cheap, and in 2026 memory shortages have been pushing prices up. The main reasons to go local are privacy, control and independence \u2013 not saving money.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"jak-model-bezi\">How a model runs on your computer<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">When you send a model a question, two phases happen, and each one is limited by something different:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>Reading the prompt (prefill).<\/strong> The model processes your entire input at once \u2013 the question, a pasted document, your project&#8217;s code. This phase depends on the GPU&#8217;s <strong>compute power<\/strong>, and it determines how long you wait for the first word of the answer. For a short question it takes a fraction of a second; for a document tens of thousands of words long it can take minutes.<\/li>\n\n\n<li><strong>Writing the answer (decode).<\/strong> The model generates text one word at a time (strictly speaking, one token at a time). For every token it has to read the whole model from memory \u2013 or, for mixture-of-experts models, just the active part. This phase depends on <strong>memory bandwidth<\/strong>: how many gigabytes per second the chip can read.<\/li>\n\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1341\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-scaled.webp\" alt=\"Reading the prompt and writing the answer: two different bottlenecks\" class=\"wp-image-29338\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-300x157.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-1024x536.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-768x402.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-1536x804.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-2048x1073.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.en_-610x319.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Example: gpt-oss-120b in llama.cpp. NVIDIA leads at reading the prompt, Apple at writing.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The practical difference: <strong>for chat<\/strong>, fast output is what you feel (and the Mac shines here). <strong>For coding agents and document work<\/strong>, which keep sending the model huge amounts of context, fast prompt reading matters just as much (and NVIDIA leads there). Speed is measured in tokens per second \u2013 one token is roughly three quarters of a word. Around 15 tokens per second matches reading speed; above 50 the model feels snappy.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"kolik-pameti\">How much memory you need<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The whole model has to sit in fast memory that the GPU can access: either the graphics card&#8217;s VRAM, or the shared (unified) memory found in Macs, the DGX Spark or the Ryzen AI Max. If it doesn&#8217;t fit, part of it spills into regular system RAM and speed collapses. In one test on an RTX 5060 Ti, moving just a sixth of the layers of Qwen 14B off the GPU cut output speed from <strong>43 to 12.5 tokens per second<\/strong>; on the CPU alone it managed 2.9.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1890\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-scaled.webp\" alt=\"Size of popular models at 4 bits compared with device memory\" class=\"wp-image-29339\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-300x221.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-1024x756.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-768x567.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-1536x1134.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-2048x1512.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.en_-610x450.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">GGUF file sizes from Hugging Face. Add memory for the context on top.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Model<\/th><th>Type<\/th><th>Size (4-bit)<\/th><th>Where it fits<\/th><\/tr><\/thead><tbody><tr><td>Llama 3.1 8B \/ Qwen3 8B<\/td><td>dense<\/td><td>4.9\u20135.0&nbsp;GB<\/td><td>anything with 8&nbsp;GB<\/td><\/tr><tr><td>Qwen3 14B<\/td><td>dense<\/td><td>9.0&nbsp;GB<\/td><td>12\u201316&nbsp;GB<\/td><\/tr><tr><td>gpt-oss-20b<\/td><td>MoE<\/td><td>12.1&nbsp;GB (MXFP4)<\/td><td>16&nbsp;GB<\/td><\/tr><tr><td>Gemma 3 27B<\/td><td>dense<\/td><td>16.6&nbsp;GB<\/td><td>24&nbsp;GB<\/td><\/tr><tr><td>Qwen3 30B-A3B<\/td><td>MoE, 3B active<\/td><td>18.6&nbsp;GB<\/td><td>24&nbsp;GB<\/td><\/tr><tr><td>Qwen3 32B<\/td><td>dense<\/td><td>19.8&nbsp;GB<\/td><td>24\u201332&nbsp;GB<\/td><\/tr><tr><td>Llama 3.3 70B<\/td><td>dense<\/td><td>42.5&nbsp;GB<\/td><td>48&nbsp;GB or more<\/td><\/tr><tr><td>gpt-oss-120b<\/td><td>MoE<\/td><td>63.4&nbsp;GB (MXFP4)<\/td><td>96\u2013128&nbsp;GB<\/td><\/tr><tr><td>GLM-4.5-Air 106B<\/td><td>MoE<\/td><td>73.0&nbsp;GB<\/td><td>96&nbsp;GB or more<\/td><\/tr><tr><td>Qwen3.5&nbsp;122B-A10B<\/td><td>MoE<\/td><td>76.5&nbsp;GB<\/td><td>96\u2013128&nbsp;GB<\/td><\/tr><tr><td>Qwen3&nbsp;235B-A22B<\/td><td>MoE<\/td><td>142.2&nbsp;GB<\/td><td>192\u2013256&nbsp;GB<\/td><\/tr><tr><td>DeepSeek R1&nbsp;671B<\/td><td>MoE<\/td><td>404.9&nbsp;GB<\/td><td>512&nbsp;GB<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"context-needs-memory-too\">Context needs memory too<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">On top of the model itself you need memory for the context (the KV cache) \u2013 everything the model &#8220;remembers&#8221; from the conversation or document. For Llama 3.3 70B that is about 10&nbsp;GB extra at a 32,000-token context and around 40&nbsp;GB at 128,000 tokens. The gpt-oss models are far more frugal thanks to a sliding window: gpt-oss-120b needs at most about 9&nbsp;GB even for 128,000 tokens. Output speed also drops as the context grows \u2013 gpt-oss-120b on an M5 Max writes about 80 tokens per second with a short context, but only 14 at 128,000 tokens.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"not-all-memory-goes-to-the-gpu\">Not all memory goes to the GPU<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>macOS<\/strong> lets the GPU use roughly 75% of RAM by default (Apple doesn&#8217;t publish the exact figure). A 64&nbsp;GB Mac therefore offers about 48&nbsp;GB for the model, a 128&nbsp;GB Mac about 96&nbsp;GB. You can raise the limit with <code>sudo sysctl iogpu.wired_limit_mb=\u2026<\/code>, but it resets after a reboot.<\/li>\n\n\n<li><strong>Windows on Ryzen AI Max<\/strong> splits memory in the BIOS at a fixed ratio \u2013 the GPU gets at most 96&nbsp;GB out of 128&nbsp;GB. On Linux it can go up to about 124&nbsp;GB.<\/li>\n\n\n<li><strong>NVIDIA RTX Spark laptops<\/strong> default to roughly a 50:50 split on Windows, with at most 96&nbsp;GB of 128&nbsp;GB set manually. The 32&nbsp;GB version gives the GPU only 16&nbsp;GB.<\/li>\n\n\n<li><strong>DGX Spark<\/strong> on Linux allocates memory dynamically, so the model can use most of the 128&nbsp;GB.<\/li>\n\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"kvantizace-moe\">Quantization and MoE, simply explained<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Quantization<\/strong> is compression for models: instead of 16 bits per parameter you use 8 or 4. At 4 bits (labelled Q4, Q4_K_M or MXFP4) a model takes up about a quarter of its original size and quality drops only slightly \u2013 a sensible default for home use. Eight bits (Q8) is nearly lossless but twice as large.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Mixture of experts (MoE)<\/strong> is an architecture where the model has a lot of parameters but only uses a small slice of them for each word. gpt-oss-120b has 117 billion parameters but activates only about five billion per step; Qwen3 30B-A3B just three billion. Such a model needs a lot of memory but writes at the speed of a much smaller one. That makes MoE models ideal for machines with large but slower shared memory \u2013 DGX Spark, Ryzen AI Max or Macs. Dense models (Llama 70B, Qwen3 32B), on the other hand, read the entire model for every word and crawl on slow memory.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"propustnost\">Memory bandwidth: why a Mac writes faster than a Spark<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">You can estimate output speed quite well: memory bandwidth divided by how much data the model has to read per token. A dense 70B model at 4 bits (~42&nbsp;GB) therefore can&#8217;t write faster than about 6\u20137 tokens per second on a DGX Spark with 273&nbsp;GB\/s \u2013 and measurements show around 4.4.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"2074\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-scaled.webp\" alt=\"Memory bandwidth of local AI hardware\" class=\"wp-image-29391\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-300x243.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-1024x829.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-768x622.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-1536x1244.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-2048x1659.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.en_-1-610x494.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Memory bandwidth as stated by the manufacturers. Under each name is the memory capacity the model can use.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"2016\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-scaled.webp\" alt=\"Local AI hardware map: memory capacity versus speed\" class=\"wp-image-29341\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-300x236.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-1024x806.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-768x605.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-1536x1210.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-2048x1613.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.en_-610x480.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Devices further right fit larger models; devices higher up write faster.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The map makes it clear that there is no perfect device. NVIDIA graphics cards have the fastest memory but little capacity. DGX Spark, Ryzen AI Max and RTX Spark have plenty of capacity but slower memory. The Mac Studio with M5 Max \u2013 and above all M5 Ultra \u2013 combines both, at the cost of no CUDA and slower prompt reading than NVIDIA.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"srovnani-gpt-oss\">Same model, different hardware<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The fairest comparison uses the same model in the same program. We picked OpenAI&#8217;s gpt-oss-120b (63&nbsp;GB) and measurements from llama.cpp, the most widely used program for running local models.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1754\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-scaled.webp\" alt=\"gpt-oss-120b speed on different hardware\" class=\"wp-image-29342\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-300x206.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-1024x701.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-768x526.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-1536x1052.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-2048x1403.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.en_-610x418.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Output and prompt reading in llama.cpp. The M5 Max and Ryzen AI Max figures are community measurements.<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"srovnani\">All devices compared<\/h2>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Device<\/th><th>Memory for AI<\/th><th>Bandwidth<\/th><th>gpt-oss-120b, output<\/th><\/tr><\/thead><tbody><tr><td><a href=\"#mac-studio-m5-max\">Mac Studio M5 Max<\/a><\/td><td>~96&nbsp;GB of 128&nbsp;GB<\/td><td>614&nbsp;GB\/s<\/td><td>105 t\/s\u00b9<\/td><\/tr><tr><td><a href=\"#dgx-spark\">NVIDIA DGX Spark<\/a><\/td><td>128&nbsp;GB<\/td><td>273&nbsp;GB\/s<\/td><td>58.7&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#rtx-spark-notebooky\">RTX Spark laptops<\/a><\/td><td>up to 96&nbsp;GB of 128&nbsp;GB<\/td><td>up to 300&nbsp;GB\/s<\/td><td>\u2013<\/td><\/tr><tr><td><a href=\"#ryzen-ai-max-395\">Ryzen AI Max+ 395<\/a><\/td><td>96&nbsp;GB (Windows) \/ ~124&nbsp;GB (Linux)<\/td><td>256&nbsp;GB\/s<\/td><td>56.6&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#mac-studio-m5-ultra\">Mac Studio M5 Ultra<\/a><\/td><td>~192&nbsp;GB of 256&nbsp;GB<\/td><td>1,228&nbsp;GB\/s<\/td><td>\u2013<\/td><\/tr><tr><td><a href=\"#mac-mini-m5-pro\">Mac mini M5 Pro<\/a><\/td><td>~48&nbsp;GB of 64&nbsp;GB<\/td><td>307&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><tr><td><a href=\"#mac-mini-m6\">Mac mini M6<\/a><\/td><td>~24&nbsp;GB of 32&nbsp;GB<\/td><td>170&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><tr><td><a href=\"#macbook-pro-m5-max\">MacBook Pro M5 Max<\/a><\/td><td>~96&nbsp;GB of 128&nbsp;GB<\/td><td>614&nbsp;GB\/s<\/td><td>105 t\/s\u00b9<\/td><\/tr><tr><td><a href=\"#rtx-5090\">RTX 5090<\/a><\/td><td>32&nbsp;GB<\/td><td>1,792&nbsp;GB\/s<\/td><td>~30 t\/s\u00b2<\/td><\/tr><tr><td><a href=\"#rtx-5060-ti-16gb\">RTX 5060 Ti 16&nbsp;GB<\/a><\/td><td>16&nbsp;GB<\/td><td>448&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><tr><td><a href=\"#rtx-5070-ti-vs-5080\">RTX 5070 Ti \/ 5080<\/a><\/td><td>16&nbsp;GB<\/td><td>896 \/ 960&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><tr><td><a href=\"#rtx-pro-6000\">RTX PRO 6000 Max-Q<\/a><\/td><td>96&nbsp;GB<\/td><td>1,792&nbsp;GB\/s<\/td><td>170.6&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#gb10-klony\">GB10 from ASUS, MSI, Lenovo\u2026<\/a><\/td><td>128&nbsp;GB<\/td><td>273&nbsp;GB\/s<\/td><td>~ same as Spark<\/td><\/tr><tr><td><a href=\"#radeon-ai-pro-r9700\">Radeon AI PRO R9700<\/a><\/td><td>32&nbsp;GB<\/td><td>~640&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><tr><td><a href=\"#ai-pc-copilot\">Copilot+ laptops<\/a><\/td><td>~8&nbsp;GB of 16&nbsp;GB<\/td><td>~135&nbsp;GB\/s<\/td><td>doesn&#8217;t fit<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u00b9 Community measurement on a MacBook Pro M5 Max with 128&nbsp;GB, llama.cpp. \u00b2 Part of the model offloaded to system RAM.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-studio-m5-max\">1. Mac Studio M5 Max \u2013 the fastest output for the money<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The Mac Studio with the M5 Max chip is today&#8217;s most sensible choice for anyone who wants the fastest possible local chat or agent and doesn&#8217;t need CUDA. With 614&nbsp;GB\/s of memory bandwidth, the 128&nbsp;GB version runs gpt-oss-120b at over 100 tokens per second \u2013 almost twice as fast as a DGX Spark. The new neural accelerators in every GPU core also sped up prompt reading, but only in up-to-date software (see <a href=\"#software\">software setup<\/a>).<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 36, 48, 64 or 128&nbsp;GB \u00b7 <strong>bandwidth<\/strong> 460&nbsp;GB\/s (36&nbsp;GB) \/ 614&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 7&nbsp;W idle, 200&nbsp;W max<\/li>\n\n\n<li><strong>Ports:<\/strong> 4\u00d7 Thunderbolt 5, 10GbE<\/li>\n\n\n<li><strong>Watch out:<\/strong> the 36&nbsp;GB version has a weaker 32-core GPU and leaves only about 27&nbsp;GB for the model. From 48&nbsp;GB up you get the full chip.<\/li>\n\n\n<li><strong>US list price:<\/strong> from $2,499 (base configuration, August 2026)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> chat and agents with MoE models, a quiet home AI server, people who also edit video. <strong>Not for:<\/strong> anyone who needs CUDA, vLLM or training.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"dgx-spark\">2. NVIDIA DGX Spark \u2013 a developer workstation, not a chat box<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The GB10 chip with 128&nbsp;GB of shared memory and NVIDIA&#8217;s full software stack (CUDA, vLLM, TensorRT-LLM) in a box 15 \u00d7 15 cm. It reads prompts faster than any other small machine \u2013 2,444 tokens per second on gpt-oss-120b \u2013 but with 273&nbsp;GB\/s of bandwidth it writes more slowly than a Mac: 58.7 tokens per second, and only 4.4 on dense Llama 70B. It draws 60\u201390&nbsp;W while running a model, and two units can be linked with a QSFP cable for models of up to around 400 billion parameters.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 128&nbsp;GB LPDDR5x \u00b7 <strong>bandwidth<\/strong> 273&nbsp;GB\/s<\/li>\n\n\n<li><strong>GPU:<\/strong> Blackwell, 6,144 CUDA cores \u00b7 <strong>AI performance<\/strong> 1 PFLOP FP4 (with sparsity; roughly half in practice)<\/li>\n\n\n<li><strong>OS:<\/strong> DGX OS (Ubuntu on Arm)<\/li>\n\n\n<li><strong>US list price:<\/strong> $3,999 at launch (October 2025), $4,699 from February 2026; ServeTheHome reported a move to a &#8220;$6,950-ish&#8221; tier in October 2026<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> CUDA development, coding agents with large contexts, fine-tuning up to 70B. <strong>Not for:<\/strong> fast chat with dense 70B models, Windows and gaming, beginners.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-spark-notebooky\">3. NVIDIA RTX Spark laptops \u2013 128&nbsp;GB in a backpack<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The N1X chip is a GB10 adapted for Windows laptops, with faster memory (up to 300&nbsp;GB\/s) and a 45\u201380&nbsp;W power envelope. In The Register&#8217;s test, a laptop ran Qwen3.6 35B-A3B at 67.6 tokens per second \u2013 faster than the desktop DGX Spark. The catch is Windows: the GPU only gets a fixed part of the memory \u2013 at most 16&nbsp;GB out of 32&nbsp;GB, 48&nbsp;GB out of 64&nbsp;GB, and 96&nbsp;GB out of 128&nbsp;GB.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 32, 64 or 128&nbsp;GB \u00b7 <strong>bandwidth<\/strong> up to 300&nbsp;GB\/s<\/li>\n\n\n<li><strong>GPU:<\/strong> Blackwell, 6,144 cores (roughly a laptop RTX 5070 in performance)<\/li>\n\n\n<li><strong>OS:<\/strong> Windows 11 on Arm, CUDA from version 13.4<\/li>\n\n\n<li><strong>Watch out:<\/strong> the first units (Surface, HP) crashed in reviews after hours of full load; Linux is not officially supported<\/li>\n\n\n<li><strong>US price:<\/strong> from $2,599 (Surface Laptop Ultra with 24&nbsp;GB) to $5,899.99 (Surface with 128&nbsp;GB), October 2026<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> large models on a Windows laptop, coding agents, creative work. <strong>Not for:<\/strong> the 32&nbsp;GB version, long training runs, x86-only software.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ryzen-ai-max-395\">4. Ryzen AI Max+ 395 \u2013 the cheapest 128&nbsp;GB<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Mini PCs with the 16-core Ryzen and Radeon 8060S graphics share 128&nbsp;GB of memory. Output speed is comparable to the DGX Spark (gpt-oss-120b at 56.6 tokens per second) for roughly half the price, and they run Windows and Linux \u2013 and games. The weak spot is prompt reading: it is two to three times slower than on the Spark and drops with long contexts \u2013 to 146 tokens per second with the Vulkan driver, 371 with ROCm.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 128&nbsp;GB LPDDR5x \u00b7 <strong>bandwidth<\/strong> 256&nbsp;GB\/s (~215 measured)<\/li>\n\n\n<li><strong>For the GPU:<\/strong> 96&nbsp;GB on Windows, ~124&nbsp;GB on Linux<\/li>\n\n\n<li><strong>Software:<\/strong> Vulkan or ROCm (gfx1151 has been officially supported since 2026)<\/li>\n\n\n<li><strong>HP Z2 Mini G1a<\/strong> costs more, but it is a certified workstation with ECC memory and a three-year warranty<\/li>\n\n\n<li><strong>US price example:<\/strong> Framework Desktop with 128&nbsp;GB launched at $1,999 and costs $2,459 after the 2026 memory price increase<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> the most memory for the money, large MoE models, chat and summaries. <strong>Not for:<\/strong> agents with huge contexts, CUDA, or the ROG Flow Z13 tablet with 32\u201364&nbsp;GB for large models.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-studio-m5-ultra\">5. Mac Studio M5 Ultra \u2013 the memory king<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Up to 512&nbsp;GB of memory and 1.2&nbsp;TB\/s of bandwidth. In the MacStories review, the M5 Ultra read prompts four times faster than the previous M3 Ultra and wrote 50% faster. The 256&nbsp;GB version even fits Qwen3&nbsp;235B. Several units can be linked via Thunderbolt 5 and RDMA: according to Apple, four M5 Ultras run an 816-billion-parameter model at 36.7 tokens per second.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 96, 256 or 512&nbsp;GB (512&nbsp;GB arrives at the end of October) \u00b7 <strong>bandwidth<\/strong> 1,228&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 9&nbsp;W idle, around 300&nbsp;W while running a model<\/li>\n\n\n<li><strong>Ports:<\/strong> 6\u00d7 Thunderbolt 5<\/li>\n\n\n<li><strong>Watch out:<\/strong> memory can&#8217;t be upgraded later; in September the wait time was up to 18 weeks<\/li>\n\n\n<li><strong>US list price:<\/strong> from $5,499 (August 2026); going from 96 to 256&nbsp;GB adds $4,000<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> 100\u2013250B models, agents with long contexts, small teams. <strong>Not for:<\/strong> models up to 30B, CUDA, tight budgets.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-mini-m5-pro\">6. Mac mini M5 Pro \u2013 the sensible entry point<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The cheapest Mac that runs genuinely useful models. 307&nbsp;GB\/s of bandwidth, up to 64&nbsp;GB of memory and Thunderbolt 5. Thanks to the neural accelerators it reads prompts about three times faster than the previous M4 Pro. The 64&nbsp;GB version comfortably runs models up to 32 billion parameters; dense Llama 70B fits but only manages about 7 tokens per second, and gpt-oss-120b doesn&#8217;t fit.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 24, 48 or 64&nbsp;GB \u00b7 <strong>bandwidth<\/strong> 307&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 6&nbsp;W idle, 145&nbsp;W max \u00b7 virtually silent at idle<\/li>\n\n\n<li><strong>US list price:<\/strong> from $1,699 (base configuration, August 2026)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> 20\u201335B models, a quiet home AI server. <strong>Not for:<\/strong> gpt-oss-120b, dense 70B, the 24&nbsp;GB version.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-mini-m6\">7. Mac mini M6 \u2013 cheap, but not with 16&nbsp;GB<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The cheapest Mac with neural accelerators. It tops out at 32&nbsp;GB and 170&nbsp;GB\/s \u2013 the 16&nbsp;GB version has only 153&nbsp;GB\/s and leaves about 12&nbsp;GB for the model. Compared with the previous Mac mini with M4, it reads prompts 4.5\u00d7 faster and writes 50% faster at practically the same price, so don&#8217;t buy the old model even on sale.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 16, 24 or 32&nbsp;GB (maximum) \u00b7 <strong>bandwidth<\/strong> 153 \/ 170&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 4&nbsp;W idle, 70&nbsp;W max<\/li>\n\n\n<li><strong>Watch out:<\/strong> Thunderbolt 4 only, so no linking of multiple units<\/li>\n\n\n<li><strong>US list price:<\/strong> from $899 (base configuration, August 2026)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> 7\u201314B models, gpt-oss-20b, Qwen3 30B-A3B. <strong>Not for:<\/strong> the 16&nbsp;GB version, models of 70B and up.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"macbook-pro-m5-max\">8. MacBook Pro M5 Max \u2013 AI on the go<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The same chip as the Mac Studio, and therefore the same speed: the fastest laptop for local models. With the mid-sized Qwen3.6 35B-A3B it generates over 90 tokens per second. The 14-inch version, however, throttles under sustained load (from 96 to 42&nbsp;W), and charging is capped at 97&nbsp;W. A Mac Studio with the same chip and 128&nbsp;GB is considerably cheaper.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 36, 48, 64 or 128&nbsp;GB \u00b7 <strong>bandwidth<\/strong> 460 \/ 614&nbsp;GB\/s<\/li>\n\n\n<li><strong>Battery:<\/strong> 72.4&nbsp;Wh (14&#8243;) \/ 100&nbsp;Wh (16&#8243;)<\/li>\n\n\n<li><strong>US list price:<\/strong> from $3,599 (14&#8243;) and $3,899 (16&#8243;), March 2026; going from 64 to 128&nbsp;GB adds $1,000<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> people who need large models on the road, ideally the 16-inch with 64\u2013128&nbsp;GB. <strong>Not for:<\/strong> the 36&nbsp;GB version, round-the-clock workloads, anyone who works at a desk.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5090\">9. RTX 5090 \u2013 the fastest, but only 32&nbsp;GB<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">1,792&nbsp;GB\/s of bandwidth: gpt-oss-20b writes at almost 300 tokens per second, and the card reads a long prompt about twice as fast as a Mac Studio with M5 Ultra. 32&nbsp;GB holds models of up to about 32 billion parameters at 4 bits, even with a long context; a 70-billion-parameter model doesn&#8217;t fit. Large MoE models can be partly offloaded to system RAM \u2013 gpt-oss-120b then runs at about 30 tokens per second.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 32&nbsp;GB GDDR7 \u00b7 <strong>bandwidth<\/strong> 1,792&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 575&nbsp;W, 1,000&nbsp;W power supply recommended<\/li>\n\n\n<li><strong>US list price:<\/strong> $1,999 at launch (January 2025); street prices in 2026 are far higher because of the memory shortage<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> coding agents and RAG with models up to 32B, CUDA development, gaming. <strong>Not for:<\/strong> 70B models and larger, a quiet low-power server.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5060-ti-16gb\">10. RTX 5060 Ti 16&nbsp;GB \u2013 the cheapest sensible entry point<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The cheapest new 16&nbsp;GB graphics card. It runs gpt-oss-20b at about 92 tokens per second, Qwen3 8B at 69 and Qwen3 14B at 41 \u2013 faster than you can read. Its memory bandwidth is half that of the RTX 5070 Ti, though, so it writes at roughly half the pace, and models around 30 billion parameters don&#8217;t fit into 16&nbsp;GB.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 16&nbsp;GB GDDR7 \u00b7 <strong>bandwidth<\/strong> 448&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 180&nbsp;W, a 600&nbsp;W power supply is enough<\/li>\n\n\n<li><strong>US list price:<\/strong> $429 for the 16&nbsp;GB version at launch (April 2025)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> first serious experiments, 8\u201320B models, a budget AI PC. <strong>Not for:<\/strong> models of 27B and up.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5070-ti-vs-5080\">11. RTX 5070 Ti vs RTX 5080 \u2013 nearly identical for AI<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Both cards have 16&nbsp;GB and almost the same bandwidth (896 vs 960&nbsp;GB\/s), so they write at nearly the same speed: 182.4 vs 184.7 tokens per second in the standard llama.cpp test, and 74 vs 81 on Qwen3 14B. Exactly the same models fit into both. The 5080 is faster mainly in games \u2013 and its launch price was $250 higher ($999 vs $749).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Our advice:<\/strong> for AI, get the RTX 5070 Ti and put the money you save into a bigger SSD or more RAM. Only buy the RTX 5080 if gaming is the main use.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-pro-6000\">12. RTX PRO 6000 \u2013 96&nbsp;GB on a single card<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The only card that fits all of Llama 3.3 70B (32 tokens per second), gpt-oss-120b (170.6 tokens per second in the Max-Q version) and Qwen3.5&nbsp;122B (about 98 tokens per second). Thanks to CUDA and vLLM it can serve several users at once. You can buy it as a standalone card or inside a pre-built workstation.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 96&nbsp;GB GDDR7 with ECC \u00b7 <strong>bandwidth<\/strong> 1,792&nbsp;GB\/s<\/li>\n\n\n<li><strong>Power:<\/strong> 300&nbsp;W (Max-Q); the full 600&nbsp;W version writes about 13% faster<\/li>\n\n\n<li><strong>US list price:<\/strong> $8,565 at launch; NVIDIA later raised it to $13,250<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> companies and teams that want to run large models in-house or train them. <strong>Not for:<\/strong> individuals and hobbyists.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"gb10-klony\">13. DGX Spark from ASUS, MSI, Lenovo, Dell and Gigabyte \u2013 which one to pick<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Five other manufacturers sell the same GB10 chip as the DGX Spark. AI performance differs by only a few percent; the differences are in storage, power supply, cooling and support \u2013 and in price.<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Model<\/th><th>Storage<\/th><th>Power supply<\/th><th>Stand-out feature<\/th><\/tr><\/thead><tbody><tr><td>NVIDIA \/ PNY DGX Spark<\/td><td>4&nbsp;TB, self-encrypting<\/td><td>240&nbsp;W<\/td><td>the original<\/td><\/tr><tr><td>GIGABYTE AI TOP ATOM<\/td><td>4&nbsp;TB PCIe 5.0<\/td><td>240&nbsp;W<\/td><td>fastest SSD<\/td><\/tr><tr><td>ASUS Ascent GX10<\/td><td>1&nbsp;TB PCIe 4.0 \/ 4&nbsp;TB PCIe 5.0<\/td><td>240&nbsp;W<\/td><td>front power button<\/td><\/tr><tr><td>Lenovo ThinkStation PGX<\/td><td>1 \/ 2 \/ 4&nbsp;TB, self-encrypting<\/td><td>240&nbsp;W<\/td><td>one year of on-site service<\/td><\/tr><tr><td>Dell Pro Max with GB10<\/td><td>PCIe 4.0 (2&nbsp;TB is QLC)<\/td><td>280&nbsp;W<\/td><td>stronger PSU, hottest SSD<\/td><\/tr><tr><td>MSI EdgeXpert<\/td><td>4&nbsp;TB<\/td><td>\u2013<\/td><td>new, no independent tests yet<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Our advice:<\/strong> the best kit for sensible money is the Gigabyte AI TOP ATOM with 4&nbsp;TB of fast storage; the cheapest proven starting point is the ASUS Ascent GX10 with 1&nbsp;TB, which launched at about $3,000 in the US. If you plan to link two units, you&#8217;ll need a QSFP cable. New since 23 October 2026: a 64&nbsp;GB version of the GB10 sold only by the partner brands, from $4,999.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"radeon-ai-pro-r9700\">14. Radeon AI PRO R9700 \u2013 32&nbsp;GB for a fraction of an RTX 5090<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The cheapest new card with 32&nbsp;GB. In the same Vulkan test it writes about 40% more slowly than the RTX 5090, but it fits the same models \u2013 around 30 billion parameters, even with a longer context. gpt-oss-20b and Qwen3 30B-A3B run on it at over 150 tokens per second. Beware of the marketing: &#8220;up to five times faster than the RTX 5080&#8221; only applies to a model that doesn&#8217;t fit into 16&nbsp;GB.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Memory:<\/strong> 32&nbsp;GB GDDR6 \u00b7 <strong>bandwidth<\/strong> ~640&nbsp;GB\/s \u00b7 <strong>power<\/strong> 300&nbsp;W<\/li>\n\n\n<li><strong>Software:<\/strong> Vulkan (usually faster output) or ROCm (faster prompt reading); no CUDA<\/li>\n\n\n<li><strong>US list price:<\/strong> $1,299 (October 2025)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> 27\u201332B models with long contexts, several cards in one PC. <strong>Not for:<\/strong> CUDA-only tools, maximum speed.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ai-pc-copilot\">15. &#8220;AI PCs&#8221; and Copilot+ laptops \u2013 a myth<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Copilot+ is Microsoft&#8217;s certification for Windows features \u2013 camera effects, live captions, Recall or the small built-in Phi Silica model. It says nothing about the ability to run large language models. Neither Ollama nor LM Studio uses the neural processing unit (NPU), and llama.cpp only experimentally; the NPU is used only by vendor tools such as Microsoft Foundry Local or AMD Lemonade with selected models.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1259\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-scaled.webp\" alt=\"Output speed of an 8-billion-parameter model: Copilot+ laptop vs graphics card\" class=\"wp-image-29343\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-300x148.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-1024x504.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-768x378.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-1536x756.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-2048x1007.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.en_-610x300.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Even a cheap graphics card is several times faster than an NPU for language models.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">On top of that, these laptops have memory with around 135&nbsp;GB\/s of bandwidth and 16&nbsp;GB of RAM, which leaves about 8&nbsp;GB for the model. They&#8217;re fine for models up to 4 billion parameters and Windows features \u2013 not for &#8220;your own ChatGPT at home&#8221;.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Good for:<\/strong> office work, battery life, small 1\u20134B models. <strong>Not for:<\/strong> anyone who wants to run 14B models or larger.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"spotreba\">Power and running costs<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">If the computer is going to run as a home AI server around the clock, power draw and noise matter too.<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Device<\/th><th>Idle<\/th><th>Running a model \/ maximum<\/th><th>Note<\/th><\/tr><\/thead><tbody><tr><td>Mac mini M6<\/td><td>4&nbsp;W<\/td><td>70&nbsp;W max<\/td><td>virtually silent<\/td><\/tr><tr><td>Mac mini M5 Pro<\/td><td>6&nbsp;W<\/td><td>145&nbsp;W max<\/td><td>5&nbsp;dBA at idle (Apple)<\/td><\/tr><tr><td>Mac Studio M5 Max<\/td><td>7&nbsp;W<\/td><td>200&nbsp;W max<\/td><td>quiet<\/td><\/tr><tr><td>Mac Studio M5 Ultra<\/td><td>9&nbsp;W<\/td><td>~300&nbsp;W running AI<\/td><td>quiet, according to reviews<\/td><\/tr><tr><td>DGX Spark<\/td><td>25\u201345&nbsp;W<\/td><td>60\u201390&nbsp;W running AI, peaks up to 200&nbsp;W<\/td><td>depends on firmware version<\/td><\/tr><tr><td>Ryzen AI Max+ 395 mini PC<\/td><td>~13&nbsp;W<\/td><td>120\u2013160&nbsp;W<\/td><td>HP Z2 Mini is loud under load<\/td><\/tr><tr><td>RTX 5090 (card only)<\/td><td>\u2013<\/td><td>up to 575&nbsp;W<\/td><td>plus the rest of the PC<\/td><\/tr><tr><td>RTX 5060 Ti (card only)<\/td><td>\u2013<\/td><td>up to 180&nbsp;W<\/td><td><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The figures come from manufacturers and reviews (ServeTheHome, Tom&#8217;s Hardware, Notebookcheck) and were not measured with the same method \u2013 treat them as a guide.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"software\">Software: what to run the model with<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Ollama<\/strong> \u2013 the easiest start on every platform; you download a model with one command. On the Mac, from version 0.19 (still in preview, for selected models) it uses MLX and the M5 neural accelerators.<\/li>\n\n\n<li><strong>LM Studio<\/strong> \u2013 a clear app with a graphical interface. <strong>Tip for M5 Macs:<\/strong> with GGUF models it doesn&#8217;t use the neural accelerators yet and reads prompts at about half the speed; use models in MLX format instead.<\/li>\n\n\n<li><strong>llama.cpp<\/strong> \u2013 the foundation of most other tools, with the most settings and the fastest support for new hardware. On Ryzen AI Max, try both backends: Vulkan usually writes faster, ROCm reads long prompts faster.<\/li>\n\n\n<li><strong>MLX<\/strong> \u2013 Apple&#8217;s library, with the best performance on the Mac, including linking several Macs over Thunderbolt 5.<\/li>\n\n\n<li><strong>vLLM, SGLang, TensorRT-LLM<\/strong> \u2013 server tools for NVIDIA when one model serves several users at once.<\/li>\n\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"chyby\">The most common buying mistakes<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>A 36&nbsp;GB Mac &#8220;for AI&#8221;.<\/strong> It has a weaker chip and leaves only about 27&nbsp;GB for the model.<\/li>\n\n\n<li><strong>An RTX Spark laptop with 32&nbsp;GB.<\/strong> Windows gives the GPU 16&nbsp;GB at most.<\/li>\n\n\n<li><strong>A Mac mini M6 with 16&nbsp;GB.<\/strong> Slower memory and only about 12&nbsp;GB for the model.<\/li>\n\n\n<li><strong>A Copilot+ laptop for local AI.<\/strong> Everyday tools don&#8217;t use the NPU.<\/li>\n\n\n<li><strong>An RTX 5080 instead of a 5070 Ti just for AI.<\/strong> The same 16&nbsp;GB, almost the same speed.<\/li>\n\n\n<li><strong>Ignoring memory bandwidth.<\/strong> Large but slow memory means slow output with dense models.<\/li>\n\n\n<li><strong>Skimping on memory.<\/strong> On Macs, Ryzen AI Max and RTX Spark the memory is soldered, and you can&#8217;t add more later.<\/li>\n\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"faq\">Frequently asked questions<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Is 16&nbsp;GB enough for local AI?<\/strong> For models up to 14\u201320 billion parameters, yes \u2013 gpt-oss-20b or Qwen3 14B, for example. Models around 30 billion need 24\u201332&nbsp;GB, and 70 billion needs at least 48&nbsp;GB.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Mac or NVIDIA?<\/strong> A Mac writes faster and gives you more memory for the money; NVIDIA reads prompts faster and has CUDA. For chat and agents with MoE models the Mac comes out ahead; for development, training and serving several users, NVIDIA does.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>What does &#8220;MoE&#8221; mean?<\/strong> Mixture of experts: the model has many parameters but uses only a small part of them for each word. It needs a lot of memory but writes at the speed of a smaller model.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>How fast does a model need to write?<\/strong> About 15 tokens per second matches reading speed. For smooth chat, 30 or more feels comfortable; for coding agents that generate a lot of text, 50 or more is better.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Is local AI cheaper than a subscription?<\/strong> Usually not \u2013 models comparable to the best cloud services need machines costing many thousands of dollars. The main reasons are privacy, control over your data and independence.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Can I link several computers?<\/strong> Yes. Two DGX Sparks connect with a QSFP cable, Macs with Thunderbolt 5 over RDMA (using MLX or exo). The Mac mini M6 can&#8217;t do this because it only has Thunderbolt 4.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"zdroje\">How we compared, and sources<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">We did not test the devices ourselves. The numbers come from verified sources, and we always cite them in the charts. We compare speed in a single program (llama.cpp), because a different program or quantization can easily shift the results by 20\u201350%. Community measurements (GitHub, forums) are marked. Prices are US list prices taken from manufacturer announcements and reviews, with their date.<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li>Model spilling into RAM, RTX 5060 Ti: <a href=\"https:\/\/inventivehq.com\/blog\/vram-offload-cliff-gpu-layers-benchmark\" rel=\"noopener\" target=\"_blank\">inventivehq.com\/blog\/vram-offload-cliff-gpu-layers-benchmark<\/a><\/li>\n\n\n<li>DGX Spark benchmark: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/benches\/dgx-spark\/dgx-spark.md\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/blob\/master\/benches\/dgx-spark\/dgx-spark.md<\/a><\/li>\n\n\n<li>gpt-oss guide and GPU measurements: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/discussions\/15396\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/discussions\/15396<\/a><\/li>\n\n\n<li>M5 Max and LM Studio: <a href=\"https:\/\/github.com\/lmstudio-ai\/lmstudio-bug-tracker\/issues\/2040\" rel=\"noopener\" target=\"_blank\">github.com\/lmstudio-ai\/lmstudio-bug-tracker\/issues\/2040<\/a><\/li>\n\n\n<li>Ryzen AI Max (Strix Halo): <a href=\"https:\/\/strix-halo-toolboxes.com\" rel=\"noopener\" target=\"_blank\">strix-halo-toolboxes.com<\/a>, <a href=\"https:\/\/llm-tracker.info\/_TOORG\/Strix-Halo\" rel=\"noopener\" target=\"_blank\">llm-tracker.info\/_TOORG\/Strix-Halo<\/a><\/li>\n\n\n<li>Apple: <a href=\"https:\/\/apple.com\/mac-studio\/specs\" rel=\"noopener\" target=\"_blank\">apple.com\/mac-studio\/specs<\/a>, <a href=\"https:\/\/apple.com\/mac-mini\/specs\" rel=\"noopener\" target=\"_blank\">apple.com\/mac-mini\/specs<\/a>, <a href=\"https:\/\/machinelearning.apple.com\/research\/mlx-distributed-inference\" rel=\"noopener\" target=\"_blank\">machinelearning.apple.com\/research\/mlx-distributed-inference<\/a><\/li>\n\n\n<li>M5 Ultra: <a href=\"https:\/\/macstories.net\/stories\/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents\" rel=\"noopener\" target=\"_blank\">macstories.net\/stories\/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents<\/a><\/li>\n\n\n<li>RTX Spark: <a href=\"https:\/\/theregister.com\" rel=\"noopener\" target=\"_blank\">theregister.com<\/a> (16 Oct 2026), <a href=\"https:\/\/servethehome.com\" rel=\"noopener\" target=\"_blank\">servethehome.com<\/a>, <a href=\"https:\/\/storagereview.com\" rel=\"noopener\" target=\"_blank\">storagereview.com<\/a><\/li>\n\n\n<li>GeForce graphics cards: <a href=\"https:\/\/nvidia.com\/en-us\/geforce\/graphics-cards\/compare\" rel=\"noopener\" target=\"_blank\">nvidia.com\/en-us\/geforce\/graphics-cards\/compare<\/a>, <a href=\"https:\/\/hardware-corner.net\/gpu-llm-benchmarks\" rel=\"noopener\" target=\"_blank\">hardware-corner.net\/gpu-llm-benchmarks<\/a><\/li>\n\n\n<li>llama.cpp leaderboards: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/discussions\/15013\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/discussions\/15013<\/a> and \/10879<\/li>\n\n\n<li>Copilot+ PC: <a href=\"https:\/\/microsoft.com\/windows\/windows-11-specifications\" rel=\"noopener\" target=\"_blank\">microsoft.com\/windows\/windows-11-specifications<\/a>, <a href=\"https:\/\/docs.ollama.com\/gpu\" rel=\"noopener\" target=\"_blank\">docs.ollama.com\/gpu<\/a>, <a href=\"https:\/\/fastflowlm.com\" rel=\"noopener\" target=\"_blank\">fastflowlm.com<\/a><\/li>\n\n\n<li>Model sizes: <a href=\"https:\/\/huggingface.co\" rel=\"noopener\" target=\"_blank\">huggingface.co<\/a> (ggml-org, unsloth and bartowski repositories)<\/li>\n\n\n<li>Prices: Apple Newsroom, NVIDIA Newsroom and developer forums, <a href=\"https:\/\/servethehome.com\" rel=\"noopener\" target=\"_blank\">servethehome.com<\/a>, <a href=\"https:\/\/storagereview.com\" rel=\"noopener\" target=\"_blank\">storagereview.com<\/a>, <a href=\"https:\/\/notebookcheck.net\" rel=\"noopener\" target=\"_blank\">notebookcheck.net<\/a>, <a href=\"https:\/\/tomshardware.com\" rel=\"noopener\" target=\"_blank\">tomshardware.com<\/a><\/li>\n\n<\/ol>\n\n\n\n<div class=\"kk-star-ratings kksr-auto kksr-align-center kksr-valign-bottom\"\n    data-payload='{&quot;align&quot;:&quot;center&quot;,&quot;id&quot;:&quot;29345&quot;,&quot;slug&quot;:&quot;default&quot;,&quot;valign&quot;:&quot;bottom&quot;,&quot;ignore&quot;:&quot;&quot;,&quot;reference&quot;:&quot;auto&quot;,&quot;class&quot;:&quot;&quot;,&quot;count&quot;:&quot;0&quot;,&quot;legendonly&quot;:&quot;&quot;,&quot;readonly&quot;:&quot;&quot;,&quot;score&quot;:&quot;0&quot;,&quot;starsonly&quot;:&quot;&quot;,&quot;best&quot;:&quot;5&quot;,&quot;gap&quot;:&quot;5&quot;,&quot;greet&quot;:&quot;Rate this post&quot;,&quot;legend&quot;:&quot;0\\\/5 - (0 votes)&quot;,&quot;size&quot;:&quot;24&quot;,&quot;title&quot;:&quot;Local AI hardware in 2026: how to choose a computer for running LLMs at home&quot;,&quot;width&quot;:&quot;0&quot;,&quot;_legend&quot;:&quot;{score}\\\/{best} - ({count} {votes})&quot;,&quot;font_factor&quot;:&quot;1.25&quot;}'>\n            \n<div class=\"kksr-stars\">\n    \n<div class=\"kksr-stars-inactive\">\n            <div class=\"kksr-star\" data-star=\"1\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"2\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"3\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"4\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"5\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n    \n<div class=\"kksr-stars-active\" style=\"width: 0px;\">\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n<\/div>\n                \n\n<div class=\"kksr-legend\" style=\"font-size: 19.2px;\">\n            <span class=\"kksr-muted\">Rate this post<\/span>\n    <\/div>\n    <\/div>\n","protected":false},"excerpt":{"rendered":"<p>A guide to choosing hardware for running language models locally: why memory capacity and bandwidth decide, how much memory popular models need, and how the Mac Studio, DGX Spark, RTX cards and Ryzen AI Max compare.<\/p>\n","protected":false},"author":1,"featured_media":29344,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_kad_blocks_custom_css":"","_kad_blocks_head_custom_js":"","_kad_blocks_body_custom_js":"","_kad_blocks_footer_custom_js":"","_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","_wrc_toc":"on","_wrc_buttons_enabled":false,"_wrc_buttons":[],"hide-affiliate-disclaimer":"","_wrc_hide_banners":"","footnotes":""},"categories":[113,173],"tags":[],"class_list":["post-29345","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-artificial-intelligence","category-guides"],"taxonomy_info":{"category":[{"value":113,"label":"Artificial intelligence"},{"value":173,"label":"Guides"}]},"featured_image_src_large":["https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/B3-18617-1024x585.png",1024,585,true],"author_info":{"display_name":"Tom\u00e1\u0161 Mahr\u00edk","author_link":"https:\/\/www.websiteradar.net\/en\/author\/maho125\/"},"comment_info":0,"category_info":[{"term_id":113,"name":"Artificial intelligence","slug":"artificial-intelligence","term_group":0,"term_taxonomy_id":113,"taxonomy":"category","description":"","parent":0,"count":20,"filter":"raw","cat_ID":113,"category_count":20,"category_description":"","cat_name":"Artificial intelligence","category_nicename":"artificial-intelligence","category_parent":0},{"term_id":173,"name":"Guides","slug":"guides","term_group":0,"term_taxonomy_id":173,"taxonomy":"category","description":"","parent":0,"count":20,"filter":"raw","cat_ID":173,"category_count":20,"category_description":"","cat_name":"Guides","category_nicename":"guides","category_parent":0}],"tag_info":false,"_links":{"self":[{"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/posts\/29345","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/comments?post=29345"}],"version-history":[{"count":1,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/posts\/29345\/revisions"}],"predecessor-version":[{"id":29392,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/posts\/29345\/revisions\/29392"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/media\/29344"}],"wp:attachment":[{"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/media?parent=29345"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/categories?post=29345"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.websiteradar.net\/en\/wp-json\/wp\/v2\/tags?post=29345"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}