🇬🇧 English · Home · Deutsche Version

Fleet Navigator runs on consumer hardware – no data centre required

Market overview: hardware for local LLM inference

As of 26 August 2026 · Price sources: Geizhals.de, Apple Store Germany, manufacturers‘ direct sales

All prices are snapshots. The memory market is currently moving week by week – always check again before you order.

What the numbers really mean

Bandwidth determines tokens per second, not TOPS

During token generation, the entire model has to be read from memory once for every single token. That puts the theoretical upper limit at roughly: tokens/s ≈ memory bandwidth / model size in bytes

A 70B model in Q4 quantisation takes up around 40 GB. On Strix Halo with 256 GB/s, that works out at about 6 tokens per second in theory. In practice, reviewers measure around 10 to 14 tokens per second on 70B models. The discrepancy arises because mixture-of-experts models do not touch all of their weights for every token. On the Mac Studio M5 Ultra with around 1.2 TB/s, you get four to five times that. This is the real difference between the three price classes, not the advertised NPU performance.

Prefill is the second, often underestimated factor

When a long context is being read in, what counts is compute, not bandwidth. This is exactly where the DGX Spark justifies its premium, and the RTX PRO 6000 with 24,064 CUDA cores and 1,792 GB/s is in a league of its own. For RAG scenarios with large document contexts, prefill performance matters more than raw generation speed. If you load many chunks into the context, you will notice the difference in time to first token far more than in output speed.

96 GB usable, not 128 GB

On all Strix Halo boxes, about 96 of the 128 GB can be addressed as GPU memory under Linux. The rest is reserved for the system. Apple’s split is more generous. If you plan with the full 128 GB as your usable model budget, your sums will not add up.

The AMD Halo family

AMD’s own reference box is called the Ryzen AI Halo Developer Platform: Ryzen AI Max+ 395 (16 Zen 5 cores, 32 threads), Radeon 8060S iGPU with 40 compute units (RDNA 3.5), 128 GB LPDDR5X-8000 unified, 2 TB SSD, 10GbE, Wi-Fi 7, 120 W system power. Officially $3,999.99 – so far a Micro Center exclusive in the USA, with no German offers.

What you can actually buy are the OEM variants with identical silicon: GMKtec EVO-X2, Framework Desktop, Beelink GTR9 Pro, HP Z2 Mini G1a. Since they all use the same chip, they differ only in cooling, networking, repairability and support.

The more powerful configuration with Ryzen AI Max+ 495, 192 GB RAM and 160 GB VRAM arrives in the third quarter of 2026. If you need 192 GB, it is worth the wait. If 128 GB is enough, there is no reason to wait, because the Max+ 395 stays in the range unchanged.

Xiaomi: two stories that keep getting mixed up

Star Core Super AI Computer: Listed on Xiaomi’s crowdfunding platform Youpin, but made not by Xiaomi but by the Chinese manufacturer Linglong. Also a Ryzen AI Max+ 395 with 128 GB, from CNY 13,999 – so no chip of its own, just another Strix Halo enclosure. An international launch is not planned.

Xiaomi AI Cube: Unveiled on 24 August 2026 at Xiaomi’s technology conference for Xring chips – a prototype with three in-house chips and a total power draw of 150 W:

Our assessment: The O100 and D100 will not reach the market until 2027, and there is no price or date for the AI Cube. So it is not relevant for a purchasing decision this year. For watching the market, it is: if China builds its own memory architecture with 1.22 TB/s, NVIDIA’s pricing structure will come under pressure in the medium term.

Why launch prices and today’s prices are so far apart

The cause is the DRAM and NAND price crisis. The driver is the DRAM contract market, not manufacturer margins. For SSDs, price levels in August 2026 are on average 126 per cent above where they stood in mid-September 2025.

What this means for purchasing: If you need 128 GB of unified memory and can afford the price, you are better off buying now than hoping for prices to fall. Forecasts for the rest of 2026 point upwards.

Decision guide

Our recommendation: Mac Mini M4 or M5 Pro

For most practices and law firms, the Mac Mini Pro with 64 GB of unified memory is the ideal solution:

• Price, Mac Mini M4 Pro: under €2,800 as a one-off investment – because of the 2026 memory price rally, check the current price on the day before you order
• Price, Mac Mini M5 Pro: Apple’s current model, prices on request
• Performance: Runs 30B models smoothly at about 30 tokens per second
• Unified memory: GPU and CPU share the same memory, ideal for large language models
• Power consumption: 60 to 80 watts, far less than a GPU server
• Noise: Whisper-quiet, fits on any desk

You keep working at your usual workstation

You do not have to work on the Mac Mini itself. Fleet Navigator routes inference across your local network to other computers. You use your Windows PC, your laptop or your office Mac as usual, while the AI computes in the background on the Mac Mini.

For this you need a Fleet Navigator Enterprise licence on the Mac Mini and a Fleet Navigator Maat licence on each additional workstation. That way several workstations share the same local AI instance, and the data never leaves your office.

It works just the same with Linux machines and multi-GPU workstations as the inference server. Here too, the workstations need a Fleet Navigator Maat licence and can use the routed inference on several other computers. This also applies to Linux machines with multiple GPUs.

What do 7B, 12B, 30B and 70B mean?

The number stands for the parameters of the AI model – more parameters = more knowledge, but more hardware required

Configurations for every requirement – our recommendations for your hardware

From entry level to a professional multi-user solution

High end: maximum performance for 70B models

For very large models, several concurrent users and maximum speed