Research Dossier ID-05 · Candidate · Updated 22 October 2026

Lucebox
Zero 495

A compact workstation built around two memory pools and an open-source engine. The useful question is what that combination lets a person finish — with what accuracy, waiting time and maintenance. The architecture is interesting; its everyday value remains a testable proposition.

RYZEN POOL 192 GB · 273 GB/S
ENGINE PLACEMENT · MERGE
R9700 POOL 32 GB · 640 GB/S
DOSSIER INSPECTOR — PLATE S1

EXPLORE THE MEMORY LAYOUT

Two distinct memory pools connected over PCIe. Capacity can be stated as a total, but bandwidth and access costs stay different. On desktop, select a node to inspect its role.

RULE: 192 + 32 = 224 GB OF MEMORY · ≠ 224 GB OF FAST VRAM
VENDOR / REPORTED product specifications, delivery estimates and published performance results. SOURCE CHECKED the public engine repository and Apache-2.0 license. UNVERIFIED BY US we have not tested a Zero 495. No independently reproduced Zero 495 result is cited in this dossier as of 22 October 2026.

Revision: engine openness confirmed; expected deliveries now differ by memory size; vendor results include the 495/192 GB configuration. The 5 October cart is retained as a historical snapshot. Sources and scope ↓

SEC.01

The spec sheet, translated.

Plate M1 — two pools, separately accounted for

Published specificationWhat to check
128/192 GB + 32 GB→Separate Ryzen unified memory and R9700 VRAM, with published bandwidths of 273 and 640 GB/s. A combined capacity figure does not describe one uniformly accessible VRAM pool. Product specifications ↗
OPEN-SOURCE ENGINE→Apache 2.0 is confirmed in the public repository. The engine supports hardware beyond Lucebox. That establishes source access; it does not prove the ease or performance of the complete appliance. License ↗
MODEL WORK ACROSS BOTH→The engine coordinates computation across the processors. Large-model profiles may include an SSD tier. Record the actual model, quantization, placement and context before comparing a result. Memory hierarchy report ↗
BROWSER SETUP + LOCAL API→The documented route uses Chrome or Edge on desktop or Android, followed by local browser management. Test the path from setup to a completed task, including recovery after an interruption. Platform ↗
340 × 110 × 320 MM · 6 KG→Approximately twelve litres: a compact workstation. Evaluate carrying it between locations separately from using it on the move. Published dimensions ↗
~40 W IDLE · ~500 W LOAD→Manufacturer estimates. At a continuous 40 W, annual energy would be approximately 350 kWh; multiply by your own tariff. Measure idle and sustained work separately. The 1,000 W PSU rating is not measured wall consumption.
£5,999 / 128 GB · £7,999 / 192 GB→Published UK base configurations, before optional storage and clustering, with VAT, duty and delivery included. Expected delivery: January 2027 for 128 GB; February for 192 GB. The £2,000 memory premium needs a workload-based justification. Current FAQ ↗
VENDOR PERFORMANCE RESULTS→The DeepSeek V4.1 report now includes 495/192 GB results, while its 128 GB figures use the earlier 395 build. Its Framework comparison changes engine, quantization and prefill method. These results guide a test; they do not isolate the effect of the extra GPU. Read the conditions ↗
SEC.02

The research register.

Plate C2 — proposed experiments, not demonstrated product capabilities

Our research interest extends to local agents that can explain which facts and permissions support an action. More memory alone does not establish better reasoning. A useful experiment needs a defined workload, a correctness check and a baseline.

Proposed object in memoryQuestion to test
Model weights and expertsFor the same task and quality target, when does placement across devices improve completion time compared with the available single-device baseline?
Facts, roles and relationsCan a structured representation reduce repeated context processing while preserving every relevant constraint and exception?
Conclusions and dependenciesAfter revoking a mandate, can the system invalidate affected decisions promptly and explain why, without retaining stale permission?
Alternative planning statesCan it compare a small set of changed assumptions and return a checkable explanation of which actions remain permitted?

First useful task

Start with one supported engine profile on hardware already available. Record setup effort, manual interventions and one completed task. Test the Zero 495 appliance separately if access becomes available; another GPU cannot establish its unboxing experience. Supported targets and harnesses ↗

Publish the conditions

For a tiered-memory run, disclose model files, quantization, engine revision, placement, SSD use, context, cache state and concurrency. Report time to first token, task completion and correctness. Compare 128 and 192 GB only with those differences visible.

Preserve the exception

The README identifies PFlash as lossy prompt compression. Begin exact-retrieval and authority tests with it disabled; then compare any compressed mode on cases with buried exceptions and revoked mandates. A faster answer counts only if the relevant constraint survives. README guidance ↗

The research result we want is a useful, reproducible task with a checkable answer.PROPOSED TESTS · NO PERFORMANCE OR AUTHORITY RESULT IS CLAIMED HERE
SEC.03

Capability map, not datasheet.

Plate M3 — workload questions across two pools and one engine

CAPABILITY MAP: questions grouped by workload. The equal-width marks are visual dividers, not benchmark scores. Read the evidence label on each row.

On the Zero 495 · Ryzen pool 273 GB/s · R9700 pool 640 GB/s · Engine between

≤ 30B MODELS
Inferred
TEST BY MODEL · QUANT + CONTEXT MATTER
70B DENSE QUANT
Inferred
COMPARE PROFILES · NO GENERAL SPEED CLAIM
LARGE MoE · ASYMMETRIC
Vendor
VENDOR RESULTS · CHECK EXACT CONFIGURATION
GIANT MoE · MEMORY TIERS
Vendor
192 GB / 128 GB · RESIDENT / SSD DEPENDENCE
SYMBOLIC / GRAPH WORK
Unverified
PROPOSED TEST · NO RESULT CLAIMED
CLUSTER KIT · 2×100GbE
Vendor
+£799 IN 5 OCT CART · SCALING TO TEST
ALWAYS-ON COMPANION
Inferred
MEASURE · IDLE, NOISE + USEFUL WORK

THE CART, AS SAVED — 05 OCTOBER 2026 · AN OFFER, NOT A RECEIPT

Zero 495 · base · 128 GB + R9700 32 GB + SSD 2 TB£5,999
Step to 192 GB+ £2,000
SSD 4 TB+ £250
Cluster kit · Mellanox CX556M · 2× 100 GbE+ £799
TOTAL IN CART£9,048

HISTORICAL SNAPSHOT · 5 OCTOBER CONFIGURATOR INCLUDED TAXES, DUTIES AND DELIVERY AND DISPLAYED A JANUARY ESTIMATE. NOT AN ORDER OR RECEIPT. CURRENT VENDOR ESTIMATE: JANUARY 2027 / 128 GB; FEBRUARY 2027 / 192 GB. THE £9,048 TOTAL INCLUDES OPTIONAL STORAGE AND CLUSTER KIT.

SEC.04

Eight tests. Claims vs evidence.

Plate O4 — proposed acceptance checks; no hands-on result claimed

T-01KettlePROMISINGTime from first setup to a useful completed task. Count manual fixes and test recovery after a connection interruption. The documented browser flow is a starting point, not a measured result.
T-02ReadingTESTUse the same documents and correctness criteria across configurations. Disclose context length, cache and compression settings. Report complete-task time alongside generation speed.
T-03LibraryMEASUREMeasure noise at a stated distance, with room background and workload recorded, including sustained operation. Fan RPM alone does not establish acoustic comfort.
T-04BackpackLIMITEDAt a published 6 kg and approximately twelve litres, treat this as a workstation that can be moved between locations. It is a poor match for routine train-seat use.
T-05Edinburgh trainCHECK OFFLINEThe vendor describes offline inference once a model is loaded. Test the whole chosen workflow with external connectivity removed; the client, tools and optional fallback may have separate dependencies.
T-06Meeting roomCHECK CONTROLTest which app or user can reach the model and which actions an agent may perform. API access and action authority are distinct questions. Verify revocation in the proposed workflow.
T-07Electricity billMEASUREMeasure at the wall: idle, active sessions and overnight operation. Report energy per completed task and actual operating hours; calculate cost with the owner's tariff.
T-08Next doorCHECK RECOVERYTest a model update, interrupted download and recovery path. Public engine code improves continuity; access to source does not by itself establish ongoing appliance support. Documented update policy ↗
SEC.05

The honest edges, up front.

Plate K5 — what the evidence can and cannot establish

prosumer@index — lucebox-zero.log
# research candidate · source review · no hands-on result $ ./lucebox-zero-495 --evidence-status [POOLS] Separate memory pools; combined capacity is not uniform VRAM. [SOURCE] Inference engine and Apache-2.0 license are public. [RESULTS] Vendor reports include multiple hardware generations. Match each number to its machine, model and conditions. [PRICE] Historical configured cart: £9,048, including options. The £2,000 memory step must earn its cost in the intended task. [WAIT] Expected delivery: Jan 2027 / 128 GB; Feb 2027 / 192 GB. A schedule is not evidence of customer shipment. [OPEN] Independent task results, sustained reliability and support. A disclosed review unit can provide evidence; customer experience adds a different and valuable perspective. $ next useful result? > One repeatable workflow, disclosed conditions, > checked outputs and an honest account of the setup effort. $
The architecture earns interest. Useful work earns a recommendation.KEEP THE TWO JUDGEMENTS SEPARATE
SEC.06

The verdict, in prosumer.

Plate Y6 — filed with interest, promoted on evidence

Consider it if

You have a defined local-model workload that benefits from the proposed memory layout, can wait for the relevant delivery window, and value an integrated setup. Establish whether the 192 GB premium improves that workload enough to justify its cost.

Skip it if

You need a machine immediately, routine backpack portability, or a measured return that current evidence cannot establish. If the workload already fits your GPU, compare the completed task on that hardware before paying for a new system.

The one test

Does the owner finish useful work reliably, with little maintenance? Measure the path from setup through a completed task, update and recovery. A convincing result should disclose the exact configuration and include the failures encountered.

CPU / POOL 1AI MAX+ PRO 495 · 128/192 GB · 273 GB/S
GPU / POOL 2RADEON AI PRO R9700 · 32 GB · 640 GB/S
CHASSIS~12 L · 340×110×320 MM · 6 KG
POWER (CLAIMED)~500 W LOAD · 40 W IDLE · 1 KW PSU
PRICE / STATUS£5,999 / 128 GB · £7,999 / 192 GB · EXPECTED JAN / FEB 2027

CLOSING — RESEARCH RULE №1 · PLATE F

Useful memory.
Checkable work.

We want to understand what heterogeneous hardware adds to a real workflow: faster completion, more usable context, predictable recovery or better control over local data.

Our authority-and-evidence experiments are proposals. They need software, test cases and a baseline; they are not capabilities established by a large memory specification.

The next step is a small, reproducible test whose result changes a practical decision.

SOURCES

Evidence register.

Reviewed 22 October 2026 · primary sources

This is a document and source review. No Zero 495 was tested for this revision. Source availability confirms what is published; it does not independently validate performance claims. Product pages and delivery estimates can change.

The cart section is the author's 5 October 2026 snapshot, retained for historical context. Corrections change the assessment when the evidence changes.