Index Dossier ID-03 · Candidate Entry · Not lived-with
A headline that writes itself — 192 GB RAM · 160 GB VRAM · 24 TB — and a spec sheet that needs a translator. This dossier doesn't review the box; it files the candidate: what the silicon is, what the headline inflates, and why the index counts machines it would never buy.
Every 192 GB machine is a negotiation between the cup and the straw. Hover a node to read the fine print on this one's terms.
Plate O1 — headline vs receipt, overprinted
| In the headline | What it actually says | |
|---|---|---|
| “192 GB RAM + 160 GB VRAM” | → | One pool of LPDDR5X, not two products. Up to 160 GB can be allocated to the GPU. Allocation is an invitation, not a velocity — the straw decides how fast the invited models drink. |
| “24 TB STORAGE” | → | The ceiling of expansion, not the contents: three M.2 2280 slots, ~24 TB max, and the described config ships a 2 TB SSD. The headline counts the wardrobe, not the clothes. Classic inflation — exactly what the index exists to deflate. |
| RYZEN AI MAX+ PRO 495 | → | 16 Zen 5 cores, 32 threads, up to 5.2 GHz — the "Gorgon Halo" platform, same silicon family as the Framework Desktop 192 GB tier. The CPU is not the bottleneck here; the memory bus is. |
| RADEON 8065S · 40 CU | → | RDNA 3.5 integrated graphics, genuinely capable — ROCm-adjacent ecosystem rather than CUDA heaven. Framework's own teardown of the 495 notes roughly 6.7% theoretical bandwidth growth over the 128 GB tier. Capacity went up 50%; the straw barely moved. |
| UP TO 160 W TDP | → | Serious heat for a mini-PC — this is workstation-class budget in a small box. Fine on a desk; a library desk at minute twenty remains an unsigned measurement. |
| DUAL USB4 · DUAL 2.5GbE | → | The connectivity story is quietly correct: fast networking for "a server you reach", USB4 for the dock life. Sane I/O, no theatre. |
| FRONT RGB ACCENT | → | The manifesto is unambiguous on this point: too old for RGB. One accent strip won't kill the deal, but it does tell you which shelf this box was designed for. |
| LPDDR5X · SOLDERED | → | Unified memory means unified forever. The RAM upgrade path does not exist — the only lever left is the three storage slots. Buy the cup you'll need in year two, at year-one prices. Ouch: that's the saffron talking. |
STATUS: CANDIDATE. THE SOURCE JD.COM CARD DID NOT RENDER READABLE SPEC TABLES — PARAMETERS ABOVE ARE AS REPORTED BY NOTEBOOKCHECK / LOWYAT / FRAMEWORK, NOT AS VERIFIED ON THE LISTING. PROMOTION TO CONFIRMED REQUIRES A READABLE RECEIPT.
Plate C2 — the anti-wishlist principle
A regional, pricey, China-first box is easy to dismiss — and dismissing it is how an index quietly becomes a list of our favourite devices. The Peladn earns its file because it marks the frontier: where the 192 GB tier is going, and at what price.
The 495's real news is the platform's direction: same CPU/GPU architecture as before, capacity as the headline, bandwidth as the footnote. Framework says it plainly — 192 GB changes what fits, not how fast it flows. That's the whole review in one sentence.
“192 GB + 160 GB VRAM + 24 TB” is three true statements that add up to a lie. The reader needs to know what's shared, what's installed and what's merely possible. This dossier is that translation, filed for the next machine that tries the same trick.
Plate Y3 — desk estimates, candidate status, straw-limited
Estimates from platform behaviour, not signed runs. The shape of this table is decided by a single fact: an iGPU drinks 192 GB through a bandwidth straw. Everything fits. Not everything flows.
CAPABILITY MAP STATUS: hypothesis / claim translation. Use the evidence label on each row before reading the adjective.
The one thing a 24 GB GPU can't do is keep a giant model resident beside your desktop, your files and your VM. 160 GB allocatable means the biggest open MoEs simply live on the machine — slow to sip, always home. For research-hoard workloads, that's the whole pitch.
If your question is "how fast does it answer", a mid-size dedicated GPU wins at half the price. The 495 answers a different question: "how much can it hold while it answers." Confuse the two and you'll return a perfectly good machine for the wrong reason.
Three M.2 slots, ~24 TB ceiling — this part of the headline is real, just not installed. For a local model archive, a document corpus, or a video hoard that must never touch a subscription, the expansion path is the most prosumer-friendly sentence on the page.
Plate K4 — candidate entry: reasoning, not receipts. TBD until signed runs.
Plate K5 — fine print, blown up to full sheet
Plate Y6 — who it's for, who it's not, filed not favoured
You run research-hoard workloads: giant MoEs that must stay resident beside your data, plus a storage ceiling that swallows archives. You accept bandwidth patience as the price of capacity, and you're inside the region — or fluent in import theatre.
You answer to tok/s rather than GB. You want CUDA's toolchain, an upgradeable cup, a quiet library desk, or a warranty in your own timezone at a sane price. None of these are character flaws. They're just other squares of the map.
Does it help me do more, or give me more to maintain? On paper the YO2 495 does one thing brilliantly — holding enormous models at home — and everything else at the speed of its straw. For the right desk that's a superpower. For most, it's a very expensive library card.
CLOSING — INDEX RULE №1 · PLATE F
The Peladn YO2 495 earns its file not despite being regional, pricey and
headline-inflated — but because of it. It maps where the 192 GB tier is going,
what the straw still refuses to do, and how easily "24 TB" becomes a feeling
rather than a fact.
Filed as candidate. Promoted only on a readable receipt.
That's not pedantry — that's the whole point of the index.