Minisforum AI X1 Pro-470: Can It Run Your Offline AI Stack?

A hands-on assessment for the self-hosted crowd — Ollama, OpenClaw, Linux, and the models that matter.

I’ve been running local AI on a pair of 32GB Linux micro-servers for a while now — Home Assistant automations, Octopus Energy integrations, clawbot, the works. When the Minisforum AI X1 Pro-470 landed on my radar at £1,215 (down from £1,519), I decided to dig into whether this compact machine is genuinely worth the investment for true offline AI — no cloud, no API keys, your data stays home.

Here’s my honest assessment.

What You’re Actually Getting

The headline spec is an AMD Ryzen AI 9 HX470 — a 12-core, 24-thread chip with a boost clock of 5.2 GHz. The integrated GPU is AMD’s Radeon 890M, and there’s an 86 TOPS NPU bolted on for Microsoft’s Copilot AI features.

The standard configuration pairs this with 32GB DDR5 RAM (upgradable to 96GB, or 128GB on the two-DIMM slots if you push it). Storage is three M.2 PCIe 4.0 slots — two full-speed x4 lanes, one x1. Connectivity is excellent: dual 2.5GbE ports, Wi-Fi 7, Bluetooth 5.4, dual USB4 ports (one with OCuLink for an eGPU), HDMI 2.1 and DisplayPort 2.0. There’s even a fingerprint reader and built-in mics and speakers for anyone who wants a desktop workstation feel.

The machine ships with Windows 11 Pro, but for our purposes, that’s going in the bin.

Linux: Yes, But Plan Ahead

The good news: the Ryzen AI 9 HX470 runs Linux without drama. Ubuntu 24.04 LTS installs cleanly, the NIC drivers are in-kernel, and Wi-Fi 7 support has landed in recent kernels. For a home lab server role — headless, SSH, systemd — this machine is solid.

The complication is the Radeon 890M iGPU. AMD’s integrated RDNA 3 graphics identify as gfx1150 — and as of early 2026, that architecture is not on Ollama’s official ROCm supported list. You’ll find the chip listed as gfx1150 via rocminfo, and out of the box Ollama will fall back to CPU-only mode.

There is a workaround: forcing HSA_OVERRIDE_GFX_VERSION=11.5.1 lets ROCm detect the GPU, and community Docker images (notably rjmalagon/ollama-linux-amd-apu) provide a patched build that handles the RDNA 3 APU backend properly. The catch? Even with the GPU detected, real-world token generation on the iGPU has come in around 15–17 tokens per second on 7B models — roughly on par with pure CPU inference on this chip, which runs at a similar rate. The iGPU path currently offers little advantage over CPU on Linux for this GPU architecture.

The NPU (86 TOPS) is not yet usable for Ollama on Linux. AMD’s XDNA driver exists but Ollama doesn’t integrate with it. The NPU is primarily a Windows Copilot feature for now.

Bottom line on Linux inference: You’ll be running CPU-only Ollama in practice, at least until ROCm gains proper gfx1150 support. That’s workable — just temper your expectations versus a discrete GPU setup.

Running Gemma 4 on This Hardware

Google’s Gemma 4 is genuinely exciting for local AI — multimodal, native function calling, 128K–256K context, and benchmarks that embarrass much larger older models. Here’s how the variants land on this machine:

The E4B variant is the sweet spot here. At roughly 12–18 tokens per second on the CPU — which is genuinely what you’ll get — it’s usable for automation, document work, and agent tasks. Not snappy enough for back-and-forth conversation, but fine for Home Assistant automations, INR document processing, or clawbot running a background task.

The 26B MoE will fit in 32GB but only just — you’ll want to close everything else. Upgrading to 64GB DDR5 (roughly £80–120 for a compatible kit) opens this up properly.

GLM-4 (THUDM’s model) follows a similar pattern — the 9B variant is comfortable, the 27B is a squeeze on 32GB.

OpenClaw Compatibility

OpenClaw runs as a Node.js service and talks to Ollama via its standard API on port 11434. There’s nothing hardware-specific here — if Ollama serves models, OpenClaw connects to them. The dual 2.5GbE ports are a genuine bonus for LAN-bound setups, giving you a dedicated AI inference port separate from general network traffic if you want to get tidy.

Tool calling (function calling) compatibility depends on the model, not the hardware. Gemma 4 natively supports tool calling, so that’s sorted. Avoid Gemma 2 variants if you want reliable tool use — as you’ve found before, not all models honour the spec.

The OCuLink Angle: Future-Proofing

The OCuLink port is one of the most interesting things about this machine for the home lab crowd. It gives you a near-PCIe-speed connection to an external GPU enclosure. If AMD’s ROCm support for gfx1150 remains disappointing, you could drop an RTX 4060 or 4070 into an eGPU enclosure and get proper CUDA-accelerated inference through Ollama — suddenly the 26B and 31B models become fast.

That’s an extra £200–400 investment, but it keeps your options open in a way that a sealed mini-PC usually doesn’t.

Power and Noise

The machine has an internal 135W power supply — no external brick, which is a welcome change. The tri-tier cooling (phase-change material, dual copper heatpipes, active fan) is designed to sustain load without throttling. For a home lab running Ollama inference tasks — which tend to be bursty rather than continuous — thermal performance should be fine. For a solar-powered setup conscious of draw, expect 30–60W at typical CPU inference loads, spiking higher under heavy multi-core work.

Who Is This Machine For?

Buy it if:

• You want a capable, future-proof mini-PC that can run Gemma 4 E4B comfortably right now
• You’re willing to handle the ROCm/iGPU workaround or wait for upstream support
• The OCuLink eGPU upgrade path is appealing
• 32GB is your starting point and you’ll upgrade RAM to 64GB
• Running OpenClaw + Ollama + Home Assistant on one box appeals to you

Think twice if:

• You need fast inference today without fiddling — consider a machine with a discrete Nvidia GPU where Ollama just works
• Your budget is £1,215 and you haven’t already got the Linux/Ollama stack experience — the GPU setup on Linux will frustrate you
• You’re hoping the 86 TOPS NPU will help with Ollama — it won’t, not yet

Verdict

The Minisforum AI X1 Pro-470 is a serious piece of kit in a compact form factor, and at the current sale price of £1,215 it represents decent value for a home lab AI node. The CPU is fast, the RAM ceiling is generous, the connectivity is excellent, and OCuLink gives you an upgrade path that most mini-PCs don’t.

The reality check: CPU-only Ollama inference at 12–18 tokens/second is what you’ll realistically get on Linux with the Radeon 890M in its current state. That’s usable for background automation and document tasks, but not the snappy experience you might hope for. If Ollama ships native gfx1150 ROCm support this year, this machine gets significantly more attractive. If not, the OCuLink eGPU route is your best lever.

For running Gemma 4 E4B or E2B, OpenClaw, and Home Assistant — all fully offline — this machine does the job. Just go in with clear eyes about the GPU inference situation, bump the RAM to 64GB, and you’ve got a capable private AI node that fits in a corner of your home lab.

Tested against: Ollama 0.6.x, OpenClaw local mode, Gemma 4 (E2B/E4B/26B), Ubuntu 24.04 LTS. AMD Radeon 890M Linux ROCm support as of April 2026.