The hardware: Strix Halo in a NAS case
All three Lucy configurations (Pro, Max and Ultra) use AMD's Ryzen AI Max+ 395, the chip codenamed Strix Halo. The review unit was the Max, with 128GB of LPDDR5X memory, a 1TB NVMe system drive, two M.2 NVMe slots and five 3.5-inch drive bays. It all sits in a Jonsbo N2 case. Jonsbo's own specifications list five 3.5-inch bays plus one 2.5-inch SSD position, Mini-ITX motherboard support, a single built-in 120mm fan, and a footprint of roughly 222.5mm on each side.
AMD's published specifications explain why this chip keeps turning up in local AI boxes. AMD's Ryzen AI Halo developer platform, which uses the same processor, lists 16 cores, 32 threads, "Zen 5" Architecture and AMD Radeon™ 8060S Integrated Graphics 40 CUs, RDNA™ 3.5 Architecture. That platform pairs the chip with 128GB of memory, with a listed Memory Speed: 8000MT/s Memory Bandwidth: 256GB/s.
The key detail for AI work is that the GPU has no memory of its own. As an iTechguides explainer puts it, the Radeon 8060S shares the system's LPDDR5x memory with the CPU, operating system, applications, and AI software. That is why the review says models have to fit in about 100GB to leave room for the operating system. It is also why "128GB" should never be read as "128GB of VRAM." The GPU specification site Flopper notes that AMD will let you address 96 GB as video memory. The ceiling you actually get depends on the OS and configuration.
What the review unit had, and what retail units are supposed to have:
| Item | Review unit | Planned retail (per reviewer) |
|---|---|---|
| Ethernet | 1Gb/s | Dual 2.5Gb/s (Pro); 10Gb/s (Max, Ultra) |
| Operating system | Ubuntu 26.04, preconfigured | Ubuntu before Christmas; Lucy AIOS after |
| System drive | 1TB NVMe | 256GB or 1TB, depending on model |
All network-related testing therefore ran over gigabit Ethernet, so treat it as a lower bound, not a preview of the shipping 10GbE model.
Summary: The hardware matches AMD's Strix Halo specification. The review sample did not match retail networking or the planned shipping software.
The storage catch: two M.2 slots, both taken
This is where the "NAS" label gets complicated. The Max has only two M.2 slots, and both are occupied out of the box:
- One holds the NVMe system drive.
- The other holds an M.2-to-SATA adapter that connects the case's five SATA bays.
If you want a second NVMe drive for model storage, you have to give up the hard-drive bays. Campanale says that trade is entirely doable, but it undercuts the whole storage pitch.
Software is the bigger issue. Cephalon told the reviewer that Lucy will have no NAS interface, by design. The idea is that you store files over SMB or NFS and let the on-device AI handle things from there. Because his unit shipped with Ubuntu and Hermes, Campanale couldn't check how retail systems will actually do this.
If you're used to Synology's DiskStation Manager or TrueNAS dashboards, expect a gap. Five bays give you physical capacity. They don't automatically give you RAID management, drive health alerts, snapshot scheduling or a clear backup workflow. If you put Ubuntu or Windows on the box, setting all of that up is your job.
Summary: Think of Lucy as an AI workstation that happens to have drive bays, not a NAS appliance with an AI add-on.
Software: lots of tinkering, and Lucy AIOS untested
Cephalon plans a custom operating system, Lucy AIOS, which the reviewer describes as built on Omarchy. It wasn't ready for review. The unit arrived with Ubuntu 26.04 preloaded with:
- Hermes, an agent harness
- Ollama running Gemma 4
- ComfyUI with Minimax H3, for image generation
- Sunshine, for game streaming
Campanale found Hermes workable, Gemma 4 average and ComfyUI slow. He blamed the slowness on current models rather than the hardware. To run bigger models, he installed his own Ubuntu 26.04 on a 2TB drive. Hermes still gave him trouble, and switching to OpenClaw fixed it.
Whatever Lucy AIOS turns out to be, this review can't tell you whether it's any good. The company's schedule (Ubuntu for units shipping before Christmas, AIOS for later ones) comes from Cephalon. Nobody has verified it yet.
Local AI performance: one reviewer's results
The models tested were gpt-oss-120b, Qwen 3.5 35B-A3B and Qwen 3.8 Flash Next. Qwen 3.8 Flash Next became his daily driver, reaching about 44 tokens per second after tuning. He also found that the Vulkan backend beat AMD's ROCm on most models. If you're running llama.cpp-style tools on Strix Halo, try both backends before deciding which is faster on your setup.
The most interesting test was a real workflow, not a synthetic benchmark. Campanale connected T3 Code on his MacBook Pro to Lucy through OpenCode and ran his own proofreading skill, which is built to catch things like product-name spacing errors that normal spell-check misses. He compared the result with GPT-6 Sol Extra High:
- GPT-6 Sol Extra High (cloud): 9 minutes 6 seconds
- Qwen 3.8 Flash Next on Lucy: 12 minutes 45 seconds
He described the two outputs as comparable to nearly identical. That's a striking result for a box on a desk, but it's one task with one custom prompt. The review doesn't give quantization levels, context lengths, software versions or repeat runs, so don't take 44 tokens per second as a guarantee.
Coding was a different story. The coding models he tried didn't keep up with cloud services, and the largest frontier-scale coding models need far more memory than any 128GB machine has.
Summary: For writing and research-style tasks, a high-memory Strix Halo box can get surprisingly close to cloud models. For serious AI-assisted coding, it doesn't yet.
Heat and noise
Under sustained AI or development load, the reviewer saw a package draw of about 138W and temperatures of 90–93°C. He was blunt: don't expect quiet token generation.
The AMD specifications add some context. AMD lists a maximum operating temperature (Tjmax) of 100°C, so 93°C is inside spec but not by much. Spec trackers such as CPU Monkey say the chip's default TDP is 55 watts and can be configured between 45 and 120 watts. A reported 138W package figure is above that ceiling. That could mean aggressive firmware tuning, or just that package power readings and TDP are measured differently. The review doesn't say how the figure was taken, so this is our analysis, not a confirmed finding. Either way, this cube shouldn't go on a desk next to your microphone.
The Windows angle
This is the part most relevant to WindowsForum readers. Campanale now runs Windows 11 Pro on the Lucy and says it works very well. A longtime Mac user, he says he replaced his M1 Max MacBook Pro with it for everything except iOS and macOS builds. He uses it for local models, development, gaming and bulk storage across the five bays.
For gaming, he puts the Radeon 8060S at roughly RTX 4060 laptop level. AMD's own marketing aims higher: AMD claims the 8060S can deliver 14% average performance increase over the NVIDIA GeForce RTX 4070 discrete graphics found in the previous generation of the ROG Flow Z13. That claim is vendor-supplied and based on a specific laptop comparison. Results in a different chassis with different power limits will vary, and the reviewer's more modest estimate is the safer one to plan around.
If you're thinking of putting Windows on it, three practical points:
- The VRAM split matters. Because memory is shared, check how much is assigned to the GPU in firmware or AMD's software before loading large models.
- Try Vulkan builds of local inference tools, given the reviewer's results against ROCm.
- You handle storage yourself. Storage Spaces or your own backup tools take the place of the NAS interface Cephalon chose not to build.
Price: check before you back it
The pricing in the review doesn't fully add up, so read it carefully:
| Model | Kickstarter (per review) | Planned retail (per review) |
|---|---|---|
| Pro | $799 | $899 |
| Max | $2,799 | $3,999 |
| Ultra | $3,999 | $4,999 |
The spec box in the same review also says "From $999," without saying which model that refers to. Cephalon told the reviewer the first 50 units should ship about a week after the campaign ends, with the rest about three months later. As with any crowdfunded hardware, those dates are promises, not guarantees, and you should confirm current prices on the campaign itself.
The reviewer works out that at $3,999 the Max would pay for itself in about 20 months against a $200-a-month AI subscription. That's simple division. It leaves out electricity, drives (which aren't counted in that figure), and the fact that he estimates local models run three to six months behind the frontier. He also notes that comparable 128GB Strix Halo mini PCs cost about $3,500 to $3,900 when the review went live, which makes the five bays look like a cheap extra.
The verdict
Is the Lucy AI Studio Max a NAS? Not really. Is it a cloud AI replacement? Not quite. It's a 128GB Strix Halo workstation in a five-bay chassis. In one experienced reviewer's hands, it ran large local models well enough to change his mind about local AI, and it also worked as a full Windows 11 Pro machine.
Before backing it, wait to see two things that haven't been tested: Lucy AIOS, and a retail unit with 10GbE. If you're comfortable setting up Ubuntu or Windows yourself, much of the value is already there. If you want a quiet, managed storage appliance, keep looking.