Computers, graphics cards and network storage that people use to run agents and local models, and to back up the machines agents run on.
12 entries. Latest source check .
What a source check covers
Documentation checked means the source was read. It does not mean the software was installed, tested or security audited.
Reading an entry does not authorise installation, credential access or spending. Tech Horizon Labs projects are labelled, and there are no affiliate links or paid placement.
Which set-up do you need?
Host the agents, rent the models
If your agents use Claude, ChatGPT, Grok or Gemini through a subscription or an API key, the box only runs the agents, their memory, files and tools, so it can be small and stay on all day. Apple Mac mini starts at US$899 with a 155 W maximum and runs macOS. GMKtec NucBox K13 (from US$589.99) and Beelink EQi12 Pro (OpenClaw pre-installed) (US$589) are x86 mini PCs that their makers sell for running the OpenClaw agent around the clock, and Intel rates their processors at 17 W and 15 W base power. Prices are from each maker's US or global store on 7 November 2026.
Two cautions. GMKtec's and Beelink's warranties exclude commercial use, so read them before you run a business agent on one. And check where your subscription may be used before you plan around it: the Models group sets out each provider's rules.
For scale, Google puts Gemma 4 31B at about 17.5 GB in 4-bit for the weights alone, and OpenAI says gpt-oss-120b fits a single 80 GB GPU. These boxes can run the agents too, but they cost more: GMKtec listed the EVO-X3 at US$3,799.99 and Beelink the GTR9 Pro at US$4,319 on their US stores on 7 November 2026.
Agents keep their memory and their work on disk. Synology DiskStation DS925+ and UGREEN NASync DXP4800 Plus can each be a Time Machine destination for Macs and back up PCs. UGREEN points out that RAID is not a backup, so keep a second copy somewhere else.
Windows 11 Pro; GMKtec says Linux works for OpenClaw
100 W adapter; Intel lists 17 W base, 37 W maximum turbo
From US$589.99, GMKtec US store, 7 Nov 2026
Figures come from each maker's or provider's own pages on the date the entry was checked. Prices and specifications change, so open the entry and its official source before you buy or build.
People building a Linux or Windows desktop or home server that runs local models and needs 32 GB of graphics memory on one card.
Access and use
For local AI inference and development on a desktop PC or workstation. It has 32 GB of dedicated GDDR6 memory on a 256-bit bus with 640 GB/s peak bandwidth, and ECC memory support on Linux only. AMD lists Windows 10, Windows 11 and Linux x86 64-bit support, and the ROCm 10.1.0 compatibility matrix (dated 26 August 2026) lists the R9700 for Ubuntu 26.04.1, 24.04.5 and 22.04.5, RHEL 10.2 and 9.8, and Windows 11 25H2. It needs a PCIe 5.0 x16 slot and two slots of space, takes a 12V-2x6 power connector and has a total board power of 300 W; AMD recommends at least a 750 W power supply. Cards come from board partners and AMD says configuration varies by partner design: XFX, for example, sells its version as model RX-97XPROAIY and specifies a 1000 W minimum power supply for it. Caveats: it is not a CUDA card, and AMD's HIPIFY documentation says HIP is not a complete replacement for CUDA and cannot translate all code, so confirm your software supports ROCm. AMD says these cards are not designed or validated for data centre use. XFX states a 2 year warranty on its card.
People who want a compact, always-on Mac to run coding agents and smaller local models, or several units as a cluster.
Access and use
For running coding agents, local models and agent workflows on macOS. The M6 model has a 12-core CPU and 12-core GPU with 16 GB of unified memory as standard, configurable to 24 GB or 32 GB, and 153 GB/s of memory bandwidth (170 GB/s with the larger memory). The M5 Pro model has up to an 18-core CPU and 20-core GPU, 24 GB of unified memory configurable to 48 GB or 64 GB, and 307 GB/s of memory bandwidth. Memory is chosen when you order. Apple's MLX framework is built for machine learning on Apple silicon and installs on macOS with pip install mlx; Apple's page also names Claude Code and Codex for coding, and exo and LM Studio Bionic for running agents across a cluster of Mac mini units, with Thunderbolt 5 on the M5 Pro model (up to 120 Gb/s) as the clustering link. Maximum continuous power is 155 W, and a power cord is in the box. It has 2.5 Gb Ethernet, configurable to 10 Gb, and runs macOS. Apple's US store listed it from $899 (M6) and from $1,699 (M5 Pro) on 7 November 2026. Caveat: the M6 model tops out at 32 GB of unified memory; the M5 Pro goes to 64 GB.
People who want to run large open-weight models and local agent workflows on one desktop Mac, or across several clustered over Thunderbolt 5.
Access and use
For running large language models and agent workflows locally. The M5 Max model has an 18-core CPU and 32-core GPU with 36 GB of unified memory at 460 GB/s; with the 40-core GPU it takes 48 GB, 64 GB or 128 GB at 614 GB/s. The M5 Ultra model has a 30-core CPU and 64-core GPU with 96 GB at 1.2 TB/s, configurable to 256 GB, or to 512 GB with the 36-core CPU and 80-core GPU version. Memory is chosen when you order. Apple's page names LM Studio Bionic for running LLMs locally and exo for clustering machines, and says Thunderbolt 5 (up to 120 Gb/s) handles clustered compute. Apple's MLX framework is built for machine learning on Apple silicon and installs on macOS with pip install mlx. Maximum continuous power is 480 W. It runs macOS. Apple's US store listed it from $2,499 (M5 Max) and from $5,499 (M5 Ultra) on 7 November 2026. Caveat: the 512 GB option is only offered on the top M5 Ultra configuration.
A 4.5 litre mini-ITX desktop from Framework with an AMD Ryzen AI Max+ 395, Radeon 8060S graphics and up to 128 GB of soldered LPDDR5x memory, sold for running large AI models on hardware you own.
People who want a small x86 desktop running Linux or Windows that can hold large quantised open-weight models in memory and run agents locally.
Access and use
For running large open-weight models and local agents on an x86 desktop. The Ryzen AI Max+ 395 versions have a 16-core, 32-thread CPU, Radeon 8060S graphics with 40 graphics cores, and 64 GB or 128 GB of soldered, non-upgradeable LPDDR5x-8000 memory on a 256-bit bus. Framework's own table shows DeepSeek V4 Flash at UD-IQ2_XXS quantisation using 90.9 GB on the 128 GB model, and notes that results vary with quantisation, context length and runtime. A newer version with the Ryzen AI Max+ PRO 495 and 192 GB is also sold. AMD's ROCm 10.1.0 matrix lists the Ryzen AI Max+ 395 (Radeon 8060S) for Ubuntu 26.04.1 and 24.04.5 and for Windows 11 25H2. Framework publishes its own list of supported Linux distributions, with Linux kernel 6.11 as the minimum and 6.15 or later recommended. The Max+ 395 is sold as a DIY Edition: you bring your own operating system (Windows 11 is an optional extra) and can bring your own M.2 2280 NVMe drives for its two PCIe 4.0 slots. It has an included 400 W FlexATX power supply, and the processor runs at 120 W sustained and 140 W boost. USB4 and a PCIe x4 slot are offered for clustering several units. Caveat: check that your Linux distribution is on both Framework's supported list and AMD's ROCm matrix before relying on GPU compute. A 1 year limited warranty is included, extendable to 3 years.
A small desktop AI computer from NVIDIA built on the GB10 Grace Blackwell superchip, with 128 GB of unified memory for running and fine-tuning large models and agents locally.
Developers who want a CUDA machine on their desk to run, test and fine-tune models of up to about 200 billion parameters and to host always-on agents.
Access and use
For running, testing and fine-tuning large models and always-on agents locally. It has 128 GB of LPDDR5x unified memory at 273 GB/s, a 20-core Arm CPU and a Blackwell GPU, and NVIDIA says it supports models up to 200 billion parameters, or 405 billion with two units. A 64 GB configuration is sold only through participating OEM partners. It ships with NVIDIA DGX OS, which is based on Ubuntu, with CUDA, cuDNN, Docker and the NVIDIA Container Runtime preinstalled; use of the preinstalled CUDA Toolkit is subject to NVIDIA's CUDA Toolkit End User License Agreement. Power comes from the included 240 W external supply (the GB10 chip is rated at 140 W), and NVIDIA warns that a different or lower-rated supply can cause reduced performance, failure to boot or unexpected shutdowns. Networking is 10 GbE, two ConnectX-7 QSFP ports and Wi-Fi 7, and up to four units can be clustered. You can use it with a monitor and keyboard or over the local network with SSH, NVIDIA Sync or remote desktop tools. NVIDIA's US marketplace listed the 4 TB model at $6,950.00, marked out of stock, on 7 November 2026; NVIDIA also sells it through authorised channel and retail partners. Caveats: the processor is ARM64-based; GPU passthrough to virtual machines is not supported; major DGX OS updates come about twice a year for the first two years after initial release, with security patches between releases.
A desktop graphics card from NVIDIA with 32 GB of GDDR7 memory on the Blackwell architecture, used to run local AI models with CUDA on Windows or Linux PCs.
People building a Windows or Linux PC or home server who need CUDA software support and 32 GB of graphics memory on one card.
Access and use
For local inference and AI development on a desktop PC. It has 32 GB of GDDR7 memory on a 512-bit bus with 1,792 GB/s of bandwidth, 21,760 CUDA cores, fifth-generation Tensor Cores and CUDA compute capability 12.0. NVIDIA's CUDA Toolkit supports Windows 11 and major Linux distributions including Ubuntu, RHEL, SUSE, Debian and Fedora, and NVIDIA names LM Studio and AnythingLLM as tools for running models on RTX PCs. Total graphics power is 575 W. NVIDIA lists 1000 W required system power, based on a PC with a Ryzen 9 9950X, and recommends a PCIe CEM 5.1 compliant power supply. Power comes from one 600 W PCIe Gen 5 cable or four PCIe 8-pin cables with the adapter in the box. The Founders Edition is a 2-slot, 304 mm long PCIe Gen 5 card; partner cards differ, and NVIDIA notes that small form factor readiness varies by manufacturer. NVIDIA's US page showed "Starting at $1999" on 7 November 2026 and points to regional NVIDIA sites for local pricing and partner retailers. Caveat: NVIDIA says power requirements differ by system configuration, so size the power supply for the whole build.
A four-bay network storage box (NAS) from Synology running DSM, with an AMD Ryzen V1500B, two 2.5GbE ports and two M.2 slots, used as a backup and file server.
People and small teams who want a backup target with snapshots for the Macs, PCs and servers their agents run on.
Access and use
For backups, snapshots and file sharing for the machines agents run on. It has a 4-core AMD Ryzen V1500B, 4 GB of DDR4 ECC memory (up to 32 GB), four 3.5 inch or 2.5 inch SATA bays, two M.2 2280 NVMe slots and two 2.5GbE ports, and takes up to nine drives with a DX525 expansion unit sold separately. It runs Synology DSM: a shared folder can be a Time Machine destination over SMB with Bonjour broadcast, it keeps up to 256 snapshots per shared folder, and Synology's backup tools cover PCs, file servers and virtual machines. Synology rates power at 37.91 W during access and 12.33 W in HDD hibernation, from a 100 W or 120 W adapter. Caveats: the DS925+ page points to Synology's drive policy for models from 2025, which says that with DSM 7.3 the 2025 DiskStation Plus models allow third-party HDDs and 2.5 inch SATA SSDs when creating storage pools, at the user's discretion; Synology only guarantees full functionality for drives on its compatibility list, and its page names Synology Enterprise Series SSDs for the M.2 slots; Synology will not provide complete warranty or support if non-Synology memory is added; C2 cloud storage for Hyper Backup is a paid subscription; Synology does not recommend the recycle bin or encrypted shared folders for a Time Machine destination. 3 year hardware warranty, extendable to 5 years.
A four-bay network storage box (NAS) from UGREEN with an Intel Pentium Gold 8505, 10GbE and 2.5GbE ports and two M.2 slots, used as a backup target for the computers agents run on.
People who want a local Time Machine and file backup target for the Macs and PCs that run their agents, with room for Docker containers.
Access and use
For local backups, file storage and light container work. It has a 5-core, 6-thread Intel Pentium Gold 8505, 8 GB of DDR5 memory expandable to 64 GB, four SATA bays, two M.2 NVMe slots for an SSD cache or SSD storage pool, a 128 GB system SSD, and one 10GbE and one 2.5GbE port. It runs UGREEN's UGOS, which provides a dedicated Time Machine destination over SMB and Bonjour (UGREEN advises a separate folder and quota for each Mac), a Sync & Backup app with a desktop client for backing up Windows or Mac folders, and Docker and virtual machines. It is sold without drives and accepts third-party drives from makers including Western Digital, Seagate and Toshiba; check UGREEN's compatibility list for the exact drive and memory models first. UGREEN rates it at 42.36 W during drive access and 18.12 W in drive hibernation, measured with a Seagate 2 TB hard drive. Reaching 10GbE speeds needs every part of the transfer path to support 10GbE. Caveats: RAID is not a backup, so keep a separate copy; UGOS backup packages use the .ubk format and must be restored through Sync & Backup on a compatible UGREEN NAS. 2 year warranty.
A small mini PC from Beelink with an Intel Core i5-1235U processor that ships with Ubuntu and the OpenClaw agent pre-installed, set up to use cloud models.
People who want a low-power box that runs an agent, its memory and tools around the clock while the models come from cloud APIs.
Access and use
For hosting an always-on agent that calls cloud models. Beelink sells it as an OpenClaw with cloud model access option and names GPT-4o, Claude and Gemini as models reached through their APIs; treat whichever provider you connect as receiving the agent's prompts and context. It has an Intel Core i5-1235U with 10 cores and 12 threads, up to 4.4 GHz; Intel lists 15 W processor base power and 55 W maximum turbo power, and launched the chip in the first quarter of 2022. Memory is 32 GB of DDR4 at 3200 MHz, listed as 16 GB x 2; Beelink does not say whether it is soldered. Storage is a 500 GB SSD, with two M.2 2280 PCIe 4.0 x4 slots for up to 4 TB each. It has two 1 Gb Ethernet ports (Realtek RTL8111H), Wi-Fi 6 and Bluetooth 5.2. Power comes from a built-in 85 W supply, and auto power-on and Wake-on-LAN are listed. The only operating system option listed is Ubuntu, with OpenClaw pre-installed; review the pre-installed OpenClaw set-up before connecting accounts. Beelink offers one-on-one setup help from its AI support team. Beelink's global store listed it at US$589 (32 GB, 500 GB) on 7 November 2026, shipping from its US, EU and UK warehouses to those addresses. Beelink states a three-year warranty, but its warranty policy says it applies only to personal use and not to products used for commercial purposes.
A desktop mini PC from Beelink with an AMD Ryzen AI Max+ 395 processor, 128 GB of soldered memory and two 10 Gb Ethernet ports, sold for local AI models and agents.
People who want to run large local models and agents on one box, or link several boxes over fast Ethernet, on Windows or Ubuntu.
Access and use
For running local models and agents, alone or as part of a small cluster. It has an AMD Ryzen AI Max+ 395 with 16 cores and 32 threads, up to 5.1 GHz, and Radeon 8060S graphics with 40 compute units. Memory is 128 GB of LPDDR5X at 8000 MT/s, soldered. Storage is two M.2 2280 PCIe 4.0 x4 slots for up to 16 TB (8 TB each), and the retail unit includes a 2 TB SSD. It has two 10 Gb Ethernet ports and two USB4 ports at 40 Gb/s, which Beelink says suit AI server clustering. Power comes from a built-in 230 W supply; Beelink lists a 120 W default TDP and 140 W peak, and auto power-on. The standard model ships with Windows 11 Pro. The OpenClaw edition ships with Ubuntu, or Ubuntu and Windows chosen at startup, with OpenClaw and a local Qwen model pre-installed. AMD lists Windows 11 and Ubuntu support for the processor. Beelink says it supports local deployment of 70B-parameter models and quotes about 52 tokens per second on GPT-OSS 120B. Orders are pre-orders with shipping in about 35 days. Beelink's US store listed it at US$4,319 (128 GB, 2 TB) on 7 November 2026; the OpenClaw edition was US$4,349 on Beelink's global store. Beelink states a three-year warranty, but its warranty policy says it applies only to personal use and not to products used for commercial purposes.
People who want 128 GB of memory for local models alongside their agents, on Windows or Linux rather than macOS.
Access and use
For running local models and agents on one box that stays on. It has an AMD Ryzen AI Max+ 395 with 16 cores and 32 threads, up to 5.1 GHz, and Radeon 8060S graphics with 40 compute units. Memory is 128 GB of onboard LPDDR5X at 8000 MT/s that cannot be upgraded, with up to 96 GB usable as graphics memory. Storage is a 2 TB or 4 TB PCIe 4.0 SSD, with two M.2 2280 PCIe 4.0 x4 slots for up to 16 TB in total. It has one 2.5 Gb Ethernet port, Wi-Fi 7, one USB4 port at 40 Gb/s and an OCuLink (PCIe Gen4 x4) port for an external graphics card; GMKtec says OCuLink is not hot-pluggable, so power off before connecting or disconnecting. GMKtec lists processor power modes of 54 W (silent), 85 W (balanced) and 140 W peak (performance). It ships with Windows 11 Pro, and GMKtec says it supports Ubuntu and Linux; AMD lists Windows 11 and Ubuntu support for the processor. GMKtec pre-installs its GMKtec Claw software for deploying local AI agents, says it can switch between cloud and local models; if you point it at a cloud model, treat that provider as receiving your prompts. GMKtec claims support for models up to 235B parameters. GMKtec's US store listed it at US$3,799.99 (128 GB, 2 TB) on 7 November 2026. The product page states a 1+2 year warranty: GMKtec's warranty page says orders from 1 October 2026 get one year on the whole unit plus two years' replacement cover for RAM, SSD and power adapter, and it excludes commercial use.
People who want a quiet, low-power box that runs agents around the clock while the models come from cloud APIs.
Access and use
For running an agent such as OpenClaw around the clock, with models from cloud services or small local ones. It has an Intel Core Ultra 7 256V with 8 cores and 8 threads, up to 4.8 GHz; Intel lists 17 W processor base power and 37 W maximum turbo power. Memory is 16 GB of on-package LPDDR5X at up to 8533 MT/s and cannot be upgraded. The spec table lists one 5 Gb Ethernet port, Wi-Fi 6E, Bluetooth 5.2 and two USB4 ports at 40 Gb/s (a feature panel on the same page says dual 2.5G LAN, so confirm before buying). It runs from a 20 V, 5 A (100 W) adapter. It ships with Windows 11 Pro; GMKtec's OpenClaw article says it is compatible with Linux and Windows 11 Pro, calls it a host for a 24/7 AI employee and says it stays cool and silent under 24/7 heavy loads. That article's setup guide starts by disabling the Windows firewall and installs Node.js and OpenClaw by piping remote scripts to bash: keep the firewall on and read any script before running it. GMKtec's US store listed it from US$589.99 (16 GB, 512 GB) on 7 November 2026. GMKtec markets its AI mini PCs as stable for 24/7 workloads, but its warranty excludes commercial use: the K13 page lists a one-year limited warranty, and GMKtec's warranty page says orders from 1 October 2026 get one year on the whole unit plus two years' replacement cover for RAM, SSD and power adapter.