Field Guide FG-01 / Lived with / Rev 2026.10.06
A 3.5-litre personal AI lab with serious GPU power and an operating idea that matters more than the spec sheet: the box is something you manage through the web, but the heavy work stays on the metal.
On the spec sheet → what it means on your desk
| Spec / claim | Evidence | Prosumer translation | |
|---|---|---|---|
| RTX 5090 Mobile · 24 GB GDDR7 | Vendor | → | Fast GPU, but the important sentence is the memory capacity: 24 GB writes the first boundary around model size. |
| 96 GB DDR5-5600 | Observed | → | The less glamorous headline. It gives large CPU/offload experiments, a Windows VM and normal desktop headroom somewhere to live. |
| Core Ultra 9 275HX | Vendor | → | A real host CPU, useful when the GPU runs out of seats. Overflow works; overflow is not the same as GPU speed. |
| ~19 dB idle / <39 dB load | Vendor | → | Promising for a desk. We still want a sustained measurement at minute twenty, in the actual room, not the vendor lab. |
| 320 × 197 × 55 mm · 2.15 kg | Vendor | → | Compact, yes. Backpack-friendly only if you ignore the 330 W power brick. This is movable lab gear, not a pocket companion. |
| Browser-managed OlaresOS | Observed | → | Apps, files, settings and management arrive through a web surface. That is the strongest everyday idea in the box. |
| TB5 · HDMI 2.1 · one USB-A · 2.5GbE | Vendor | → | Enough I/O, but honestly sparse. One USB-A is a lifestyle choice. The 2.5GbE port is the quieter clue: this box makes more sense as a machine you reach than one you constantly sit at. |
Same metal, two jobs, one surface
The web UI is excellent for management and review. Heavy builds, model servers and long agent sessions belong on the host. Use the door as a door.
A Windows VM can live beside the Linux host. One physical box can be both local AI infrastructure and the Windows machine you still occasionally need.
Once several Olares boxes expose the same management language, “check the other machine” starts feeling less like changing computers and more like opening another room. The fleet shrinks to one surface.
Working classification, not a signed benchmark sheet yet
No grudge. Just receipts.
At the rough edges, app state, containers and internal networking become visible. The moment I need Kubernetes vocabulary to ask my own model a question, the product has handed me work I did not order.
The GPU is quick. The 24 GB capacity determines which models remain pleasant rather than merely technically possible.
The market, firmware and polish depend on one young vendor. If Olares vanished, the hardware would still be a PC — but a stranger, more orphaned one.
After the setup evening, do I keep opening it — or does the browser chatbot win because it is easier? For this box, the uptime and night-shift use say it still earns the shelf.
Evidence status is part of the score, not a footnote
| Test | Current call | Evidence | Why |
|---|---|---|---|
| First result | Pass-ish | Observed | Onboarding and one-click apps can get to useful local AI quickly when the happy path behaves. |
| Reading | Measure | Unverified | Needs repeatable model + document run with first-token and throughput numbers. |
| Library | Edge | Vendor | Published acoustics are close to the target; sustained room measurement still owed. |
| Backpack | Fail | Vendor | Chassis plus 330 W brick makes this transportable lab gear, not the light carry profile. |
| Offline | Edge | Observed | Local models work locally; “zero outbound packets” is a separate test and remains unverified. |
| Meeting room | Test | Inferred | The architecture fits the use case; a full no-network client-document workflow still needs a real run. |
| Electricity | Measure | Unverified | Idle and sustained wall power should be measured with a plug meter. |
| Next door | Edge | Observed | Expansion and fleet paths exist, but some combinations introduce bandwidth, power or software caveats. |
The device was the purchase. The experience became the asset.
Other machines may be faster at one job or cheaper on paper. Olares One mattered because it became a place to discover what local AI is for, what breaks, and what a prosumer should demand next. That is an author field note — not a product fact.