WiCi One GPU: Maker's RTX 5090 PC Hits 7.27 tok/s, Level With DGX Spark

WiCi, maker of the WiCi One wireless GPU box, puts its RTX 5090 test PC at 7.27 tokens per second (tok/s) decoding DeepSeek-V4.1-Flash-1017 in its own open benchmark, level with NVIDIA's DGX Spark at 7.26 tok/s. The WiCi One shop page claims 20 tok/s for the $9,999 RTX 5090 edition on DeepSeek V4.1 Flash. Those two maker figures were taken under different conditions, and only the first comes with conditions stated.

That gap is what a buyer needs to understand before ordering. The "750B model on 32 GB of VRAM over Wi-Fi" pitch does not come from the radio. It comes from InferSSD, an open-source engine that streams model weights from an SSD, and WiCi's own repository shows that engine running on ordinary PCs, a DGX Spark and a MacBook Pro. No independent outlet or lab has published an LLM speed for a WiCi One. This file sets out what the company states, what its own benchmark shows, and what remains unmeasured.

What WiCi One Is and What It Costs

WiCi One is a box with an NVIDIA RTX GPU, an Intel CPU and an NVMe SSD inside. It plugs into a wall outlet and serves laptops, tablets and phones as a GPU over Wi-Fi 7, with a 4x4 MIMO radio on a 320 MHz channel, according to WiCi's shop and overview pages. Devices reach it three ways: a driver that presents a virtual GPU on macOS, Windows and Linux, an SDK for Python, JavaScript and Swift, or an OpenAI-compatible API. WiCi says the box is bought outright, with no subscription.

The shop lists two editions. The RTX 5060 Ti edition has 16 GB of GDDR7, an Intel Core 3 100U and a 1 TB NVMe 4 SSD for $2,599, and its buy box reads "Ships in 8 weeks". The RTX 5090 edition has 32 GB of GDDR7, an Intel Core Ultra 7 255H and a 4 TB NVMe 5 SSD for $9,999, and the shop gives no separate ship date for it, though its FAQ says orders ship within 8 weeks of purchase. Both pages show a $600-off code applied at checkout, which would bring the 5060 Ti edition to an effective $1,999, the early-access price that Notebookcheck (Sept. 25) and Engadget (Sept. 26) reported. WiCi does not state the system RAM of either retail box.

Gizmodo reported on Oct. 5 that the 5060 Ti box uses a 650 W desktop power supply and the 5090 box a 1,000 W unit. It also reported a Thunderbolt 4 port for running the box as a conventional external GPU. WiCi's own pages list no Thunderbolt port, so that detail rests on Gizmodo's report alone.

What's New: InferSSD Streams Weights From the SSD

The new part is the way a model larger than the GPU's memory gets to run at all. WiCi's overview page says the box runs InferSSD, which keeps the weights on the onboard NVMe and moves the parts in use to the GPU. The shop page calls the same idea "TurboStream weight paging". It works because DeepSeek-V4.1-Flash is a mixture-of-experts (MoE) model: the InferSSD README lists it at 763B parameters in total but only 8B to 16B active per token. According to the InferSSD changelog, it streams the routed experts from the SSD through bounded RAM and VRAM caches, so only the experts a token needs have to sit in fast memory.

InferSSD is public. It is licensed under Apache-2.0, is built on llama.cpp and reads standard GGUF files. Version 0.1.0 shipped on Oct. 30, and version 0.1.1, with Linux CUDA build fixes, followed on Oct. 31. It installs on Apple Silicon Macs and on Linux machines with NVIDIA GPUs, and the README says the first run downloads about 470 GiB of model files. The engine is not tied to WiCi's hardware.

The README benchmarks DeepSeek-V4.1-Flash-1017 as a Q8_0 GGUF, with routed experts kept at native 4-bit MXFP4, at a 4K prompt and 256 output tokens, with decode measured over tokens 129 to 256. Under those conditions it lists decode at 7.27 tok/s on the RTX 5090 PC, 7.26 tok/s on a DGX Spark, 2.874 tok/s on a MacBook Pro with an M5 Max and 48 GB, and 2.55 tok/s on the RTX 5060 Ti PC. Prefill runs at 164.5 tok/s on the 5090 PC and 127.5 tok/s on the DGX Spark.

Those test machines are not the boxes WiCi sells. The README's RTX 5090 PC uses an Intel Core Ultra 5 230F, 32 GB of system RAM and a 2 TB PCIe 5.0 SSD, against the retail edition's Core Ultra 7 255H and 4 TB. The RTX 5060 Ti PC shares the retail Core 3 100U but has 8 GB of system RAM and a 2 TB PCIe 4.0 SSD, against 1 TB in the retail box. The README does not mention Wi-Fi or a WiCi One unit: these are local runs on the PCs themselves. The methodology page also says the Linux and NVIDIA rates are tester-supplied summary values.

How WiCi One Compares With the DGX Spark

WiCi's argument is that one box on the home network can serve a large open model to every device in the house without a cloud subscription. Wi-Fi is not what limits the chat case. WiCi's FAQ says models load from the onboard NVMe and are not sent over the air, so a chat request moves a prompt one way and a stream of text tokens back, a small load for a Wi-Fi 7 link. The link matters far more for graphics and games, where frames cross the air. In Gizmodo's Cyberpunk 2077 demo, performance was close to 30% better next to the box than across the room.

The comparison a local-AI buyer should draw is with the DGX Spark. In WiCi's own table, the RTX 5090 test PC and the Spark decode within 0.01 tok/s of each other on the same model, quantization and window. The 5090 edition lists at $9,999. NVIDIA's marketplace listed the 128 GB DGX Spark at $6,950 on Oct. 31, with stock unverified, as covered in our DGX Spark review. The Spark also has 128 GB of unified memory and can run the same open engine. What the WiCi box adds is the wireless serving layer and the RTX GPU for graphics work, not a faster decode on this model by its maker's own numbers.

WiCi One Speed: What Independent Tests Have Not Measured

The largest open question is where the shop's speeds come from. The shop says its LLM figures were taken "over Wi-Fi 7 on a production WiCi One" and gives DeepSeek V4.1 Flash at 20 tok/s on the 5090 edition and 8 tok/s on the 5060 Ti. It states no quantization, context length, prompt size or measurement window. By Voltgrid's arithmetic, the 20 tok/s figure is 2.75 times the 7.27 tok/s in WiCi's open benchmark, and 8 tok/s is 3.14 times the 2.55 tok/s; no source states those ratios. The two sets were measured on different hardware under different or unstated conditions, so the ratio is not a like-for-like gap. WiCi has not explained the difference.

WiCi's materials also disagree on the model's size. The shop says 748B parameters, the InferSSD README says 763B in total, and the description of a sponsored video says 750 billion. The shop also lists Kimi K3, at 2.8 trillion parameters, at 2.5 tok/s on the 5060 Ti edition and 5 tok/s on the 5090 edition. InferSSD 0.1.0 does not list Kimi K3 as a supported model, and WiCi has not published how the Kimi K3 figures were measured. They remain maker claims, although the shop's comparison table calls the 5 tok/s figure measured.

No independent outlet or lab has measured LLM speed on a WiCi One. Gizmodo's Oct. 5 report, "The First 'Wireless GPU' Uses Wi-Fi to Headbutt All Those AI Subscriptions", centred on gaming from a demo with WiCi chief executive Zili Meng. The 5090 unit ran DeepSeek-V4.1 on a short prompt, but Gizmodo reported no speed, and it wrote that the product still needs work before launch. Notebookcheck and Engadget reported the prices and WiCi's own performance claims. NetworkChuck's video of Oct. 30 is marked "Sponsored by WiCi", so it counts as a maker-sponsored demo, not an independent measurement.

When Independent WiCi One Tests Are Expected

The 5060 Ti buy box read "Ships in 8 weeks" on Oct. 31, and Gizmodo reported a December launch plan; eight weeks from Oct. 31 would be late December, by Voltgrid's arithmetic. Independent LLM figures can come only after retail units reach testers. The useful run would put a retail 5090 edition and a DGX Spark through InferSSD on DeepSeek-V4.1-Flash-1017 Q8_0 at the README's settings, once over Wi-Fi and once locally. Until then, every speed quoted for WiCi One is the company's own.