Best hardware for local LLMs and AI agents in 2026

TL;DR by budget

• Under €150: Raspberry Pi 5 (8 GB) + 27W PSU + microSD + active cooler + Tailscale subscription. ~€140 all in.
• €300-500: Generic mini PC (Geekom IT13, Beelink SER7 or similar), 32 GB RAM, 1 TB NVMe. ~€450-500.
• €600-1100: Mac Mini M4 (16 GB or 24 GB), default storage. €699-1099.
• €1500-2200: Mac Mini M4 Pro (48 GB or 64 GB unified) or Framework Laptop 16. The local-LLM endgame at small form factor.
• €3000+: Mac Studio or custom workstation with discrete GPU. Diminishing returns beyond this for most personal/small-team workloads.

The body of this guide expands each tier, compares the realistic options, and lists what each tier does for self-hosted AI.

1. How this guide was put together

This is a buyer's guide, written in April 2026 and revised in October 2026, from the makers' specifications and the agents' documentation. PocketClaw has not bought or tested these devices (see /methodology). The speed figures quoted below are published measurements from the llama.cpp project and the Slips project, each with its setup; the full tables are in Local LLM benchmarks. Prices are the April 2026 figures the guide was written with; they move, so check the seller.

The one thing that sorts the tiers before anything else is memory: a local model has to fit in RAM (or in Apple's unified memory) to run at a usable speed, and an agent runtime needs room next to it. Here is the most memory each device can be ordered with, from the makers' own pages.

The scale is linear: the single-board computers are a sliver next to the Mac Studio because they top out at 16 GB, a thirty-second of the largest Mac Studio on sale.

For scale on the model side: a 7B or 8B model at 4-bit is a 4 to 5 GB download in the Ollama library, and Llama 3.3 70B is 43 GB. The model-by-memory table lines up the common sizes.

2. Tier 1: single-board computers (€80-300)

The Raspberry Pi 5 is the canonical SBC for self-hosted AI in 2026, with the Orange Pi 5 Plus as the closest credible competitor. The Khadas Edge2, Banana Pi BPI-M4 Berry and Radxa Rock 5B+ sit in the same price band; none of them has the software ecosystem of the two leaders.

2.1 Raspberry Pi 5 (8 GB): the default

The Pi 5 is the device this guide assumes you'll start with unless you have a specific reason to choose otherwise.

On paper: a Broadcom BCM2712 with four Cortex-A76 cores at 2.4 GHz, 1 to 16 GB of LPDDR4X, and, in Raspberry Pi's words, “2–3× the speed of the previous generation” (product page).

What it suits:

• An agent that calls a cloud model (Claude, GPT, OpenRouter) and runs its tools locally: the model's work happens elsewhere, so the Pi mostly waits on the network.
• Small local models. The Slips project measured about 11 tokens per second with 1B models and about 5 with 3B models on a Pi 5, with Ollama at Q4_K_M (Slips documentation). It needed 4.7 GB of memory for llama3.2:3b: take the 8 GB board.

What it does not suit:

• 7B models and up. No published Pi 5 measurement of a 7B model gives its setup, and the 3B figure above is already slow for long prompts: in the same tests, a 5,000-token prompt took a 3B model about 15 minutes.
• Agents that drive a full browser, which needs memory the 8 GB board shares with everything else.

Power and heat, from Raspberry Pi's documentation: the recommended supply is the 27 W USB-C unit, and the typical bare-board active current is 800 mA, about 4 W at 5 V, before any SSD, cooler or load (power supply). The SoC throttles its cores from 80 °C and both cores and GPU at 85 °C; the documentation notes that a heatsink or small fan reduces throttling (frequency management). For anything close to 24/7, fit the official active cooler.

Cost breakdown (April 2026):

• Pi 5 8 GB board: €85
• Official 27W USB-C PSU: €13
• microSD 64 GB Class 10: €10
• Official active cooler: €5
• Optional NVMe HAT + 256 GB SSD: +€60 (recommended for serious use)
• Total: ~€113-173 depending on storage choice

We recommend the NVMe route for any setup that will run more than a week: an agent writes logs, model files and container layers constantly, and microSD cards are not made for that kind of load.

2.2 Orange Pi 5 Plus (16 GB): the spec-pusher

The Orange Pi 5 Plus is more capable on paper than the Pi 5: an RK3588 (four Cortex-A76 and four Cortex-A55 cores) with a 6 TOPS NPU and a Mali-G610 GPU, and up to 16 GB of LPDDR4/4X on the maker's page. That headroom matters for memory-bound workloads.

The catch is software ecosystem. Official Orange Pi OS images are functional but rough. Armbian community builds are better, but you'll spend time learning a less-trodden distribution. The Pi's “everything just works” story isn't replicated here.

When we recommend the Orange Pi 5 Plus over the Pi 5:

• You specifically need 16 GB of RAM in an SBC form factor
• You're comfortable on Armbian
• The RK3588 NPU integration matters for your workload

When we don't:

• First-time SBC user
• Production where community support matters
• Any case where Pi-specific tutorials would save you time

Cost (April 2026): 16 GB SKU around €180.

2.3 The other SBCs

Briefly:

• Khadas Edge2: an RK3588S2 board with 8 GB, or 16 GB on the Pro configuration; nice hardware, smaller community than the Orange Pi. Specs on its page here.
• Radxa Rock 5B+: capable, reasonable price, but ecosystem feels sparse compared to the Pi.
• Banana Pi BPI-M4 Berry: not one we would pick for sustained agent workloads; smaller community and thinner documentation than the Pi.

If you want SBC and you're not sure, get the Pi 5. The Orange Pi 5 Plus is competitive, but only if you know what you're getting into.

3. Tier 2: mini PCs (€300-700)

The middle of the buyer's guide and where most readers should land. Mini PCs at this price point hit the sweet spot of capability, portability, power efficiency and software compatibility, and they are the usual answer to “best mini PC for a local LLM” under €1,000.

3.1 Generic mini PCs: the workhorse

Geekom IT13 (i7-13620H), Beelink SER7 (Ryzen 7 7840HS) and ASUS PN-series boxes, in 32 GB / 1 TB configurations, are the common picks for self-hosted AI workloads.

What decides local-LLM speed on these boxes is the integrated graphics and the memory behind it, not the CPU core count. Published llama.cpp measurements, generating with the same 7B model (Llama 2 7B at Q4_0) through the Vulkan backend:

• AMD Ryzen 7000-series integrated graphics: 19.91 tokens per second (the family of the Ryzen 7040 chips in boxes like the SER7; the scoreboard does not name the exact chip).
• Intel Core Ultra 100-series integrated graphics: 8.21 tokens per second.
• The scoreboard has no entry for the 13th-generation Intel Core chips of the Geekom IT13 and the Intel NUC 13.

Source: llama.cpp's Vulkan scoreboard, read on 17 October 2026; builds and the rest of the table in Local LLM benchmarks.

What doesn't fit:

• Llama 3.3 70B at Q4: the file is 43 GB, so it does not load in 32 GB. For 70B-class models, look at Apple Silicon with 64 GB or more.

Cost: €450-550 for 32 GB / 1 TB SKUs in April 2026. Ryzen options were generally cheaper than Intel at similar specs.

Our pick at this tier: a 32 GB box. If local models matter, one built on a Ryzen 7040-series chip, whose family posts the best integrated-graphics figure above; if they don't, the Geekom IT13 at €450-490 (April 2026) is a mainstream Intel option.

3.2 Intel NUC 13 / NUC 14 (now ASUS-branded): the reference

After Intel exited the NUC business in 2023, ASUS picked up manufacturing. The current NUC 13/14 line is the reference design that generic mini PCs clone.

Pros: design quality, BIOS stability, broader vendor support.

Cons: a price premium over generic equivalents. Not worth it unless you specifically value reference design (compliance, bulk procurement, etc.).

Cost: €700-1000 for equivalent 32 GB / 1 TB SKUs.

3.3 The fanless option

Several vendors (CWWK, Topton, Akasa) make fanless mini PCs based on Intel N-series or low-TDP i3 CPUs, such as CWWK's N100 boxes.

Pros: literally silent, low idle power, capable for agent workloads that call a cloud model.

Cons: 16 GB RAM ceiling on N100, single-channel memory. Local LLMs are essentially out of scope: on the llama.cpp Vulkan scoreboard, the Intel N150 writes 2.93 tokens per second with the same 7B model.

When this makes sense: noise-sensitive deployment (bedroom homelab), edge MCP server host, dedicated tool runner.

Cost: €250-350.

4. Tier 3: Apple Silicon (€699-2200)

Apple's unified memory architecture is the cheat code at this price range. RAM is also LLM working memory, and memory bandwidth is the chip's strength: a Mac mini M4 with 24 GB or more does things at this price point that x86 mini PCs in the same range cannot.

Apple has replaced both machines since this guide was written. Its spec pages, read on 16 October 2026, list the Mac mini with M6 and M5 Pro chips (up to 64 GB of unified memory) and the Mac Studio with M5 Max and M5 Ultra chips (up to 512 GB). The M4 figures below are those of the Mac mini (2024): 16 GB as standard, configurable to 24 or 32 GB at 120 GB/s, and 24 GB on the M4 Pro, configurable to 48 or 64 GB at 273 GB/s.

4.1 Mac mini M4 (16 GB / 24 GB / 32 GB): the local-LLM entry

The base M4 Mac Mini is good value at €699. The 16 GB SKU is fine for cloud-LLM-driven agent workloads. The real value starts at 24 GB (€899), which leaves room for an 8B model next to the agent runtime.

Published speed: 24.11 tokens per second generating with Llama 2 7B at Q4_0 on the M4 with the 10-core GPU, and 221.29 tokens per second reading a 512-token prompt (llama.cpp's Apple Silicon scoreboard, Metal, build of November 2023). On the same model, the AMD Ryzen 7000-series integrated graphics above write 19.91 and the Intel Core Ultra 100 series 8.21.

4.2 Mac mini M4 Pro (48 GB / 64 GB): the 70B target

The M4 Pro tier is where local-LLM self-hosting gets serious. Its memory bandwidth (273 GB/s, against 120 GB/s for the M4) roughly doubles generation speed on the same model: 49.64 tokens per second with the 16-core GPU and 50.74 with the 20-core GPU, on the same llama.cpp test.

For 70B-class models, memory decides: Llama 3.3 70B at Q4 is a 43 GB download. At 48 GB it loads with about 5 GB left for macOS, the context and the agent: tight. 64 GB is the configuration that leaves room. Neither llama.cpp scoreboard has a 70B row, so this guide gives no 70B speed.

At €1,899 for 48 GB, the M4 Pro Mac Mini was the cheapest Apple machine that could load a 70B-class model at all in 2026.

4.3 Mac mini M4: the macOS lock-in question

Asahi Linux on Apple Silicon Macs is impressive but not production-grade for server workloads as of April 2026. If your stack absolutely requires Linux, the Mac Mini story breaks down.

If you can run macOS as your host OS, the Mac Mini is the most capable small-form-factor AI host at its price point. If you can't, move to mini PCs.

5. Tier 4: Framework laptops (€1100-2200)

Framework's repairable, modular laptops earn their place in this guide because they're the rare modern laptop where Linux is first-class and upgrades are realistic.

5.1 Framework Laptop 13: the development machine

Ryzen 7 7840U or Intel Core Ultra 7 with up to 64 GB DDR5. A laptop where Linux is a first-class citizen. Repair-friendly, upgrade-friendly.

For self-hosted AI as “work laptop that also hosts an agent in the background”, the FW13 with 32 GB RAM (~€1500) is the cleanest option in 2026.

5.2 Framework Laptop 16: laptop with discrete GPU

The FW16 is more controversial. The discrete GPU module (Radeon RX 7700S) adds local-LLM headroom but trades portability and battery life.

If you want one machine for both daily-driver work AND demanding local LLMs, the FW16 is the most credible portable option. If you're willing to use a Mac Mini at home + a lighter laptop for portability, that combination is generally better and cheaper.

5.3 The general laptop-as-server caveat

Laptops as always-on servers have inherent compromises:

• Battery hardware that ages whether you use it or not
• Thermal/fan tuning optimised for portability, not sustained load
• Sleep/wake behaviour that's never quite what you want for 24/7

If your AI host is going to be on continuously, get a desktop-class device (mini PC or Mac Mini). If you have one budget for one machine that does both, Framework is the right answer.

6. Tier 5: workstations (€3000+)

Above the Mac Mini M4 Pro tier, you're moving into Mac Studio, custom workstation, or rack-mount territory. Diminishing returns set in quickly for most personal/small-team workloads.

6.1 Mac Studio M3 Ultra: the spec king

96 GB of unified memory as standard, configurable to 256 GB, with 819 GB/s of memory bandwidth and a 28-core CPU and 60-core GPU (configurable to 32 and 80), per Apple's Mac Studio (2025) tech specs read on 16 October 2026. Published speed on the same 7B test: 88.40 tokens per second with the 60-core GPU, 92.14 with the 80-core GPU. Apple has since moved the Mac Studio to M5 Max and M5 Ultra chips; the M5 Ultra can be ordered with 512 GB.

When this makes sense: deployments where local-LLM performance is the bottleneck, models too large for 64 GB, and capability that matters more than cost. Otherwise, the M4 Pro Mac Mini covers models up to the 70B class for a fraction of the price.

6.2 Custom workstation with discrete GPU

A €2,000-3,500 custom build with an RTX 4090 or RTX 5080 trades the unified-memory advantage for raw inference throughput on models that fit in the card's own memory, plus PCIe expandability and the ability to run non-Apple runtimes (vLLM, SGLang, etc.) at full speed. On the llama.cpp Vulkan scoreboard, the RTX 5080 writes 185.10 tokens per second and the RTX 4090 187.97 on the same 7B test.

For developers building production AI infrastructure, a custom workstation may be preferable to Apple Silicon. For “self-hosted personal agent” workloads, the Mac Studio remains simpler and more power-efficient.

6.3 Rack-mount and home-lab tier

Beyond the workstation tier, you're building a home lab, not a “pocket AI” setup. Out of scope for this guide.

7. Use case to recommendation

Six common cases, condensed.

8. What we would not buy (and why)

Devices we would not buy for this, and why:

• Older Raspberry Pi (Pi 4 8 GB): still works, but Raspberry Pi gives the Pi 5 2-3 times the speed of the previous generation for €30-40 more. Just buy the Pi 5.
• NVIDIA Jetson Nano / Orin Nano: AI-accelerator boards built for vision and ML projects; expensive for an agent host that mostly calls a cloud model.
• Older Apple hardware (M1 Mac Mini): refurbished M1 minis are cheap, but the memory bandwidth gap shows on local LLMs: 14.15 tokens per second for the M1 (8-core GPU, 68 GB/s) against 24.11 for the M4 on llama.cpp's Apple Silicon scoreboard, same 7B test.
• Tiny x86 sticks (Intel Compute Stick et al.): too constrained, too thermal-limited.
• Generic Chinese mini PCs without thermal review: the bottom end of the no-name market is a minefield. Stick to Geekom, Beelink, Minisforum, or branded options.

9. Hosting infrastructure beyond the device

The hardware is half the buying decision. The other half is the software/network stack that goes with it. Briefly:

• Reverse proxy: Caddy is the easiest path to HTTPS-everything with automatic Let's Encrypt. Nginx if you have specific needs Caddy doesn't cover.
• Container runtime: Docker is the default. Podman is fine. lxd/lxc for power users.
• Remote access: Tailscale. Don't expose your agent dashboard to the public internet.
• Backup: borgbackup or restic to a remote endpoint. Test the restore once a quarter.
• Monitoring: Netdata, Prometheus + Grafana, or just a status page on the agent itself.

10. Power, cooling, and physical placement

Often overlooked but matters in practice:

• Always-on hosting in a closed cabinet without ventilation is bad for everything. Add some airflow.
• Power consumption at 24/7: only some makers publish figures. Apple gives 4 W idle and 65 W maximum for the Mac mini M4 (5 W and 140 W for the M4 Pro); 4 W around the clock is 35 kWh a year, about €7 at €0.21/kWh. Raspberry Pi gives a typical bare-board current of 800 mA for the Pi 5, about 4 W, before storage, cooler and load. For a mini PC, measure at the wall with a plug-in meter: the maker's power adapter rating is a ceiling, not a consumption figure.
• UPS for production: a small APC or CyberPower UPS (€60-100) saves you from one corrupted database the moment the power blinks. Worth it.
• Wall mounting / desk placement: the Pi 5 is small enough to Velcro behind a monitor. Mini PCs can VESA-mount. Mac Minis prefer table-top.

11. The pocket-AI setup we recommend in 2026

If you want our specific recommendation for “set up once, run forever”:

1. Hardware: €450 generic mini PC, 32 GB RAM, 1 TB NVMe.
2. OS: Debian 12 minimal install.
3. Container runtime: Docker + Docker Compose.
4. Reverse proxy: Caddy (auto-HTTPS via Let's Encrypt).
5. Remote access: Tailscale.
6. Agent: Hermes Agent latest, full tool set including browser.
7. LLM stack: Claude API as primary (Anthropic via OpenRouter), Ollama with Mistral 7B Q4 as offline fallback.
8. Backup: borgbackup nightly to a Hetzner Storage Box (€3.50/month for 1 TB).
9. Monitoring: Netdata for system stats.
10. Updates: Watchtower for container updates, unattended-upgrades for Debian security patches.

This setup costs €500 hardware + €5/month recurring (Tailscale + Hetzner storage box), plus the electricity, which depends on the box: measure it. Compare with €30-50/month managed-hosting alternatives, €1,800-3,000 over five years.

12. Closing notes

The pocket-AI thesis is straightforward: hardware in 2026 is good enough that “run your AI on a device you own” is competitive with cloud-hosted alternatives for most personal and small-team use cases. The catch is that you have to be willing to do the operational work: set up the host, manage updates, monitor the system, fix things when they break.

If that work is interesting to you, you're the audience for this guide and we'd love to have you on the newsletter. If it isn't, the managed-hosting services exist for a reason: ClawRift, NitroClaw and friends will gladly take $19 to $100 a month to handle it for you, by their pricing pages read on 18 October 2026.

Either way: pick deliberately. Don't buy more than you need. Don't under-buy and ship a setup that can't grow.

Hardware in this guide

• Raspberry Pi 5: check price
• Intel NUC 13: check price
• Mac Mini M4: check price
• Orange Pi 5 Plus: check price
• Framework Laptop: sold direct by Framework, no affiliate

The check-price links are Amazon affiliate links: they cost you nothing and fund the site. Full list on /disclosure.

The hardware pages live at /pocket. The agents at /agents. Subscribe to the newsletter to hear when this guide is revised.

Related

• Local LLM benchmarks: published speeds, model by machine
• Pocket AI complete guide: which agents run on which device class
• Self-hosted AI agents 2026: landscape report
• The complete OpenClaw timeline