By Evan Vega
A YouTube video that ran a 748-billion-parameter AI model on a box with a 32 GB graphics card holds up under checking,
Novel Cognition found, though how it compares with a big-memory Mac depends on which model you run. NetworkChuck’s
30 October 2026 video, sponsored by WiCi and disclosed as such, shows WiCi One, an RTX 5090 or RTX 5060 Ti in a Wi-Fi 7
box that every laptop, tablet and Mac in the house can use as its own graphics card. On camera it ran DeepSeek V4.1 Flash
at about eight tokens a second by streaming most of the model from a solid-state drive.
Novel Cognition checked 21 of the video’s claims against DeepSeek’s model card, WiCi’s open-source InferSSD code and
benchmark notes, and Apple’s specification pages, then ran four tests on its own Mac Studio on 31 October 2026. The full
analysis, with charts and sources, is at familygpu.novcog.us.com.
NetworkChuck’s sponsored video ran DeepSeek V4.1 Flash, about 748 billion parameters, on WiCi One’s 32 GB RTX 5090 at roughly eight tokens a second by streaming the model from an SSD. Novel Cognition
WiCi lists WiCi One at $2,599 with an RTX 5060 Ti and $9,999 with the RTX 5090 the video tested, each $600 less with a
discount code. Orders ship within eight weeks to US addresses and can be cancelled for a full refund until they ship;
after that, sales are final, with a defective unit replaced if reported within seven days and a one-year warranty. Apple
lists the Mac Studio from $2,499 with an M5 Max and 36 GB, and from $5,499 with an M5 Ultra and 96 GB, configurable to 256
or 512 GB. WiCi’s own comparison table uses the 36 GB M5 Max model.
Both price lists side by side.
Of the 21 claims Novel Cognition checked, 12 hold as stated, six hold with a caveat and three could not be tested; none
failed. DeepSeek’s model card describes V4.1 Flash as a mixture-of-experts model with 384 routed experts and one shared
expert in each layer, six of them active for each token, which is why only a slice of its roughly 500 GB of weights has
to be read for each word. InferSSD, released under the Apache-2.0 licence and built on llama.cpp, packs each expert as one
aligned read, keeps the experts that recur in memory and, on Linux with an NVIDIA card, caches some on the card itself.
Every claim and its source is on the analysis site.
Three caveats matter to a buyer. “Full quality” refers to the text build the engine loads, which leaves out image input
and DeepSeek’s built-in draft model. The same model has three published speeds: 7.27 tokens a second in WiCi’s own repository
on an RTX 5090 PC with a prompt of about 4,000 tokens, about eight in the video on an early engineering unit, and 20 on WiCi’s
product page, which does not state its test conditions. And on a Mac, InferSSD computes every routed expert on the
processor rather than the graphics chip; its repository reports 2.87 tokens a second on a 48 GB MacBook Pro.
How the SSD streaming works.
Novel Cognition runs a Mac Studio with an M3 Ultra and 256 GB of memory as a shared AI server for its office, called
over the network for drafting, voice, transcription and search. Asked the video’s two coffee questions, it ran DeepSeek V4
Flash, the smaller 284-billion-parameter previous model, entirely in memory at a median 22.5 tokens a second over six
runs. That is not the same model as the video’s: V4.1 Flash scores higher on DeepSeek’s agent benchmarks, and its official
release, about 510 GB, does not fit in 256 GB.
The other tests isolate what WiCi’s design works around. The Studio’s own Thunderbolt model drive read large blocks at
2.6 GB/s and 4 KB blocks, one at a time, at 34 MB/s, 76 times slower. Switching back to the chat model after loading a
second, smaller model took 0.69 seconds because neither had to leave memory, against 17 seconds without WiCi’s scheduler
and about four with it on the 32 GB card in the video. A streamed answer used about 18 kilobits a second on the office
network, against roughly 400 megabits a second for a game over WiCi’s Wi-Fi protocol. Details:
the stack,
sharing the GPU and
Wi-Fi or API.
Novel Cognition’s reading is that the two approaches do different jobs. For the largest open models, without a
big-memory Mac, WiCi One runs them and shares its graphics card with every device in the house; the same free engine also
runs on an ordinary RTX 5090 PC. For models that fit in a Mac’s memory, a big-memory Mac Studio keeps them all loaded, with
nothing to swap and no drive in the loop. The next test is WiCi’s engine running V4.1 Flash on the 256 GB Studio. Every
source is listed at sources, and the analysis is also an episode of
the Frontier Watch podcast.
This analysis was produced under the Hidden State Drift method, which
checks every claim against its primary source before publication. The ten pre-publication checks behind it are published
at hiddenstatedrift.com/verify. Novel Cognition is not affiliated with
WiCi, NetworkChuck, DeepSeek or Apple, and nobody paid for this analysis.
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily California Press
