FRAMEWORK DESKTOP 495 — An Index Dossier

SEC.01

The spec sheet, translated.

Plate L1 — promise vs fine print, overprinted

In the brochure		What it actually says
“MODULAR · REPAIRABLE”	→	True — for the skin. Swappable front tiles, expansion cards, standard fasteners, user-serviceable everything except the board. Read on; the exception is the expensive part.
RYZEN AI MAX+ PRO 495 + 192 GB	→	Same "Gorgon Halo" platform as the Peladn — and both parts are soldered. Framework's own spec page says so. The most saffron-priced components in the box are the ones you can never swap.
“+50% MEMORY = +50% SPEED”	→	Framework's own teardown is admirably blunt: same CPU/GPU architectures, same memory bandwidth class — about 6.7% theoretical gain, plus 64 GB of capacity. The cup grew; the straw didn't. Their blog says it better than any review could.
OPEN-ENDED PCIe x4 SLOT	→	Accepts long cards like the Intel E810 for 50 GbE with RDMA — the real clustering path. But: not externally accessible in the standard case. Possible on the board ≠ convenient in the box. Plan the chassis before the card.
PRE-INSTALLED NOCTUA 120 MM	→	A genuinely encouraging sign for the Library test — Noctua is a quiet-first brand. Encouraging is not a dB(A). Minute twenty, your desk, signed.
120 W / 140 W (BOOST)	→	Careful: that's the CPU's rating — sustained and raised — not measured wall power of the whole machine. A 4.5 L system with fans, NVMe, NIC and losses will draw differently. Index rule: never copy a chip's TDP into a electricity-bill row.
FEDORA OR WINDOWS 11 · DIY / PREBUILT	→	An OS choice that respects the audience: Linux-first is native prosumer habitat. DIY or pre-built, your call — the platform is the product.
PRE-ORDERS: SEP 30 → SHIPS NOV	→	As filed, nobody outside Framework has one. Everything below the hardware row is platform inference plus community testimony from the previous Strix Halo generation. Candidate entry, like its Peladn cousin.

STATUS: CANDIDATE. SPEC ROWS FROM FRAMEWORK'S OWN SPEC SHEET AND TEARDOWN POST; CLUSTERING ROWS FROM THE OFFICIAL ANNOUNCEMENT PLUS A PUBLISHED COMMUNITY BUILD ON THE PRIOR PLATFORM. PROMOTION AWAITS SHIPPED HARDWARE AND SIGNED RUNS.

SEC.02

The clustering clinic.

Plate C2 — one word, three different proofs

"Can I cluster two?" is three questions wearing one coat. Framework itself answers them differently — USB4 for one thing, fast NICs for another. Even the manufacturer describes modes, not a universal switch.

What you want from two machines	What actually needs proving
Run different models / agents in parallel	Task distribution convenience and total throughput. The friendliest case — USB4 at ~30 Gb/s after BIOS update (from ~9) is built for exactly this: pipeline parallelism, multi-agent systems. Most achievable, least glamorous.
Run one model that doesn't fit in one box	Real model-splitting support, enough combined memory, acceptable speed. This is tensor-parallel territory: 50 GbE with RDMA via E810-class cards — real hardware path, real ROCm/RCCL labour. A published community build made a two-node Strix Halo cluster work with vLLM… and describes the networking as serious work, not a checkbox.
Make one model answer faster	Measured speedup on the specific sharding mode — not the sum of spec sheets. "Linear scaling" is Framework's claim; from "the network is slower than local memory" you cannot conclude either scaling or its impossibility. The method decides what crosses the wire; the wire decides the bill.

What exists today

A documented, working two-node Strix Halo cluster: Intel E810, RDMA/RoCE, vLLM — with substantial ROCm and RCCL effort honestly described. Important caveat: that testimony is the previous platform, not an independent test of the 495. Faith-adjacent, evidence-light — filed as such.

What remains a claim

"Linear scaling" across nodes is a manufacturer statement awaiting an independent signed run. The index doesn't mock it and doesn't invoice against it. Claims get filed in pencil; measurements get ink.

Who pays for clustering

You do — in evenings. Distributed inference in 2026 still bills in driver versions, NIC firmware and RoCE verbs. If your terminal ceiling is one pasteable block, a two-node tensor-parallel lab will find your ceiling and raise it a career.

“Not fiction, but not plug-and-forget. The cluster is real; the 'easy' is workload-dependent — and the evenings belong to ROCm.” — THE VERDICT ON 'CAN I CLUSTER TWO?', ONE SENTENCE, AS PER INDEX RULES

SEC.03

Capability map, not datasheet.

Plate L3 — platform estimates plus clustering rows, candidate status

CAPABILITY MAP STATUS: hypothesis / claim translation. Use the evidence label on each row before reading the adjective.

On the Desktop 495 · 192 GB unified · RDNA 3.5 straw · same cup-and-straw as ID-03

≤ 14B MODELS

Inferred
FLIES · SINGLE NODE

~30B CLASS

Inferred
COMFORTABLE · DAILY DRIVER

70B DENSE QUANT

Inferred
WORKS · READING SPEED

GIANT MoE · 160 GB CLASS

Inferred
RESIDENT · THE CUP'S WHOLE POINT

2-NODE · PARALLEL AGENTS

Unverified
MOST ACHIEVABLE · USB4 ~30 GB/S

2-NODE · TENSOR PARALLEL

Unverified
REAL BUT WORKFUL · 50GbE RDMA + ROCm

“LINEAR SCALING”

Vendor
VENDOR CLAIM · UNSIGNED

ALWAYS-ON COMPANION

Inferred
NO · 4.5 L DESKTOP CLASS

SEC.04

Eight tests. Honest marks only.

Plate O4 — reasoning from documentation, receipts pending shipment

T-01KettleTBDFedora-first, screw-free panels, no RGB ritual. All signs point to painless — but the box hasn't shipped. No kettle, no verdict.

T-02ReadingHYPOTHESISSame straw as ID-03: comfortable to ~30B, patient beyond. Framework's own 6.7% figure is the whole speed story. Signed runs will draw the exact line.

T-03LibraryPROMISINGA pre-installed Noctua 120 mm is the most reassuring fan sentence in this dossier series. Still: model unknown, dB(A) unpublished. Minute twenty decides.

T-04BackpackNO4.5 litres, mini-ITX, desktop-class. Category fail, not quality fail — as with every lab in this series.

T-05Edinburgh trainIN PRINCIPLEFedora, local models, no vendor cloud required on paper. Zero-packet verification awaits a shipped unit and an owner's evening.

T-06Meeting roomIN PRINCIPLEResident giant MoE behind a repairable chassis — good pitch. Signed logs and clean-offline behaviour: unverified, filed.

T-07Electricity billMEASURE120 W is the CPU's promise, not the wall's confession. Whole-system idle watts are the number the index needs. Plug-in meter, one week.

T-08Next doorSPLITTiles, cards, fan, storage: gloriously upgradeable. CPU and RAM: soldered, forever. And the PCIe x4 — the one slot that matters for clustering — lives inside a case that doesn't let it out. Choose your chassis accordingly.

SEC.05

The honest edges, up front.

Plate K5 — fine print, blown up to full sheet

prosumer@index — framework-desktop.log

# the most documented machine in the dossier series. still: candidate. $ ./framework-desktop-495 --list-honest-edges [SOLDERED] cpu + 192 gb, both bga, both forever. the repairable philosophy has a boundary, and it's drawn in gold pins. [SLOT] open-ended pcie x4 — real, long-card friendly, e810-approved... and not reachable from outside the standard case. capability ≠ convenience. house distinction, kept. [WATTS] 120 w is the processor's rating. the wall will tell a different story. never copy a chip's tdp into a tco row. [CLUSTER] usb4 ~30 gb/s (post-bios) for pipelines; 50gbe rdma for tensors. real modes, real work, real ROCm sundays. "linear scaling" stays in pencil until independently signed. [TEARDOWN] +6.7% bandwidth for +64 gb. their own blog said it. respect. the cup grew; the straw held its ground. [STATUS] pre-orders sep 30, ships november. filed as candidate, promoted on receipt — same rule as the peladn, no favourites. $ so who is this machine? > the platform vendor. sells the map, the screwdriver, > and the honest footnotes — while soldering the two treasures to the board. > you may applaud and read the spec sheet. both. in that order. $

“Framework's blog answers the speed question before any review can: the cup grew 50%, the straw grew 6.7%. That's not a criticism. It's the most honest sentence in the 192 GB marketing cycle.” — WHY THIS DOSSIER QUOTES THE MANUFACTURER AT LENGTH

SEC.06

The verdict, in prosumer.

Plate Y6 — filed, respected, neither loved nor dismissed

Consider it if

You want the 192 GB cup inside a philosophy you can open with a screwdriver, you read teardown blogs for pleasure, and you intend to cluster — with your eyes open about ROCm evenings. Fedora-first is your habitat, not your homework.

Skip it if

You answer to tok/s, not GB. You need the PCIe card reachable without case surgery. Or the soldered heart bothers you more than the repairable skin comforts you — that's a legitimate verdict, and the index files it too.

The one test

Does it help me do more, or give me more to maintain? The Desktop 495 gives you a platform, a map and a fan that respects silence — and quietly forecloses the two upgrades that matter most. On my ledger: the footnotes outnumber the friction. Rare. Noted.

PLATFORMRYZEN AI MAX+ PRO 495 · 16C/32T

GPU / RAM8065S 40 CU · 192 GB SOLDERED

CHASSIS4.5 L MINI-ITX · TILES · NOCTUA 120MM

EXPANSIONCARDS + PCIe x4 (NOT EXTERNAL IN CASE)

POWER / OS120W CPU RATING · FEDORA / WIN11

CLOSING — INDEX RULE №3 · PLATE F

Read which parts can actually leave
before you buy the philosophy.

Framework deserves the rarest compliment this index keeps: it documented its own limitations before we could. Soldered heart, buried slot, CPU watts masquerading as system watts — all printed on their own pages, in their own words.

That's not a gotcha. That's the standard every other dossier in this series will be measured against. Claims in pencil, measurements in ink — and honesty pre-installed.