WiCi One: NetworkChuck Tests a Wireless GPU for Local AI

NetworkChuck demonstrates a thin laptop playing a demanding game and running a huge AI model through a nearby WiCi One with an RTX 5090. The intriguing part is not that a 763-billion-parameter checkpoint somehow fits inside 32 GB of GPU memory. It does not. WiCi combines a GPU shared over the local network with an SSD-backed inference engine that moves selected model weights through limited memory. Those are two distinct engineering problems, and they deserve two distinct tests.

Watch NetworkChuck's Demonstration

Sponsorship: WiCi sponsored the video and supplied an early engineering unit. NetworkChuck offers a $600-off sponsor link. The link and the demonstration are not an independent hardware endorsement.

Source note, 31 October 2026: The video and supplied transcript establish what NetworkChuck saw and measured. We checked specifications, prices and policies against WiCi's product page; the engine, supported platforms and benchmark method against its public repository; and the official parameter count against DeepSeek's model card. We did not receive a device, rerun a benchmark or audit the wireless protocol.

Two Problems, Two Systems

Problem one is distance. A conventional GPU sits on a fast internal bus. Sending every driver call across Wi-Fi would add a network round trip many times during a single render or inference. WiCi says its driver and protocol reduce those blocking calls through caching, deduplication and pipelining, while keeping the GPU on your local network. The laptop remains the host for the application; the box does the GPU-heavy work. WiCi also offers an SDK and a local, OpenAI-compatible API, which are different integration routes from a virtual GPU driver. These are WiCi's architecture claims, not an independent compatibility test.

Problem two is capacity. Model weights can be far larger than a single GPU's VRAM or a typical computer's RAM. WiCi's InferSSD, built on llama.cpp, stores a supported GGUF on an SSD and streams the mixture-of-experts weights needed as generation proceeds. Active pieces are cached in faster memory. It trades storage capacity for extra data movement and lower speed; an SSD does not literally acquire VRAM's bandwidth. InferSSD is public software and can be tested without buying WiCi One.

How a 763B Model Runs With 32 GB of VRAM

DeepSeek's official model card lists 763B total parameters. The tested community GGUF conversion is not that complete original checkpoint: WiCi's format notes say it is text-only and excludes the vision and multi-token-prediction components. InferSSD's README describes about 470 GiB of first-run model files, stored on disk rather than in VRAM. The 5090 edition has 32 GB of VRAM and a 4 TB NVMe SSD, according to WiCi's specifications.

The helpful mental model is a library: the SSD holds the shelves, RAM and VRAM hold a changing selection of pages, and the mixture-of-experts router determines which parts are needed next. Most of the model is not active for every token, but the full set must remain available on storage because future tokens may choose different experts. In the video, NetworkChuck explains aligned reads, batched requests, caching popular experts and placing work across CPU and GPU. WiCi's architecture document gives the implementation details. This makes the run possible; it does not make the first download, model loading or response time disappear.

What the Speed Numbers Mean

The same model name appears beside several very different numbers. Here is the distinction to keep in mind before treating any one figure as a purchase guarantee.

The numbers are not interchangeable. The repository test used 32 GB system RAM, 32 GB VRAM and a 2 TB PCIe 5 SSD, while WiCi lists a different CPU and 4 TB SSD for the 5090 product. Its published decode rate is a late-output window, not a whole-request average or time to first token. Ask for repeatable tests with your exact model file, prompt length, cold and warm starts, concurrent users and network conditions. A faster benchmark on an optimized production unit is possible; it is not established by the public standalone result alone.

The Game, Creative Apps and Shared GPU

NetworkChuck shows Cyberpunk running on a thin laptop through the WiCi box, and he tours creative workflows including Blender and video work. At 19:41, he switches between chat and image generation: one switch drops from roughly 18 to 8.6 seconds, another from 17 to 4 seconds with WiCi's scheduler. That is a useful illustration of model parking and reuse, not a controlled multi-user throughput benchmark. WiCi says model weights can be shared while each workload keeps its own state; check isolation and capacity under your own simultaneous load.

The Figma and Lovart examples need special care. Both remain cloud applications. For this demo, their AI-generation requests were redirected to a model running locally on WiCi One; that does not make the complete application, its accounts, storage or telemetry local. Likewise, WiCi's wireless link still carries data across your network. A local inference endpoint can improve control of one data path, but privacy depends on every connected app, permission and route.

Try InferSSD Before Buying a Box

WiCi's Apache-2.0 repository currently documents Apple Silicon macOS and Linux with an NVIDIA GPU; Windows is marked in progress. It builds from source and requires platform-specific tools. Its DeepSeek example fetches about 470 GiB, so choose the SSD location before a first run. The official quick start has the current install, model-directory and serving commands. Read the installer and dependencies before executing them, and verify there is enough free storage for the model and working files.

Start by recording a normal task on a model that already fits your machine; then compare quality, time to first token, full-request time, watt-hours and total setup effort with the large GGUF. WiCi's own table shows the large DeepSeek conversion decoding at 2.550 tokens/s on an RTX 5060 Ti test rig. That proves the issue is not a strict yes/no VRAM threshold. The question is whether a large model is useful enough at the resulting speed for your work.

Before You Buy WiCi One

WiCi lists two editions as of 31 October 2026. These are vendor list prices before the advertised $600 discount, shipping and tax, not independently checked checkout totals.

The current shipping policy says orders ship within eight weeks to US addresses only. You can cancel for a full refund before shipping; after shipping, WiCi says change-of-mind sales are final, with a seven-day arrival-defect process and a one-year limited warranty. That matters especially for readers outside the United States. The video shows an early unit, so confirm retail software, supported host OS and apps, router requirements, noise, power draw, sustained performance and concurrent-device behavior before committing.

Finally, “no cloud inference bill” is not “free AI.” Hardware, electricity, cooling, SSD wear and maintenance all count. WiCi itself says the box complements rather than replaces cloud AI. If you only occasionally need a frontier-size model, a metered hosted service may be cheaper and simpler; if several trusted devices need sustained local compute, the economics may change. That is the comparison worth doing.

Resources by Role

Video Chapters

The wireless GPU

The large-model problem

Sharing, price and verdict

Sources and Limits

The video and user-supplied transcript are the source for NetworkChuck's game, creative-app and model-switching demonstrations. The product page and technical overview are WiCi's own descriptions. The InferSSD repository and benchmark methodology establish what its public release and test report actually cover. DeepSeek's checkpoint and the community conversion are separate artifacts. Pricing, availability, performance and compatibility can change. We have not independently reproduced WiCi's results or inspected the retail product.