{"slug": "wici-one", "title": "WiCi One", "summary": "WiCi One, a wireless GPU device from an unnamed company, is now available for pre-order at $1,999 USD for early signups, down from $2,599, and claims to deliver real-time AI inference and gaming performance over Wi-Fi 7 to existing devices. The device features a full-power GPU, Wi-Fi 7 radio, and a software stack supporting local AI models like Ollama and LM Studio, with benchmarks showing up to 500+ tok/s for Gemma 4 and 110 fps in Cyberpunk 2077 at 4K Ultra with DLSS 4. The product aims to provide a private, low-latency alternative to cloud computing for AI workloads.", "body_md": "# WiCi One Untethered GPU Power\n\nWiCi One wirelessly equips the devices you already own with the best GPU — real-time, private AI without making anything larger, heavier, or hotter.\n\nStarting at ~~$2,599~~ **$1,999** USDFor early signups only\n\n### Top 10 ideas get *WiCi One free.*\n\nSign up for the newsletter and a chance to get one free in the Pioneer plan — we'll send you the details.\n\n##\nLet the GPU *come to you.*\n\nAI is becoming personal. WiCi One is the wireless compute layer for this new era — one dedicated device that equips your phones, laptops, and future AI devices with the best GPU over Wi-Fi.\n\n### A wireless compute layer.\n\nCompute becomes a shared resource every device can access wirelessly — quiet, always on, always near.\n\n### Not device-only.\n\nPowerful AI shouldn't make devices larger, heavier, or hotter. Your devices stay light — WiCi One carries the load nearby.\n\n### Not cloud-only.\n\nSkip the cloud's tradeoffs — latency, ongoing costs, privacy, and a hard dependence on connectivity.\n\n## A new kind of\n\n### GPU runtime\n\nA full software stack runs on the main board — driver, runtime, scheduler, and developer tools. Everything an app needs to treat WiCi One as nearby compute.\n\n### Wi-Fi 7 radio\n\nA high-bandwidth wireless module tuned for the WiCi protocol. Twin antennas keep nearby devices in lockstep — low latency, always available, no cables.\n\n### Full-power GPU\n\nA dedicated graphics processor sized for real-time AI inference. The heavy lifting happens here, just a few feet from your phone, your laptop, your speaker.\n\n### Chassis & cooling\n\nA precision-machined enclosure with a quiet thermal path beneath a curved top. Built to sit on a shelf and be forgotten, the way infrastructure should feel.\n\n## Real workloads,\n\n*served near you.*\n\n### Works with the apps *you already use.*\n\n### AI & local models\n\n- Ollama\n- LM Studio\n- Hugging Face\n- PyTorch\n- ComfyUI\n- Open WebUI\n\n### Video & streaming\n\n- Premiere Pro\n- CapCut\n- OBS Studio\n\n### 3D & creation\n\n- Blender\n- AutoCAD\n- Rhino\n- Maya\n- Unreal Engine\n- Unity\n\n### Games\n\nA few highlights, not the full list. If an app can use a GPU, it can use WiCi One with no changes needed: Driver mode shows up as a standard local GPU, and API mode serves standard endpoints. Your apps and games run on your own machine — not remote play, just a wireless GPU. [Read our developer section ](developers)\n\nBenchmark† | 5060 Ti edition | 5090 edition |\n|---|---|---|\n| Kimi K32.8T params | 2.5 tok/s |\n5 tok/s |\n| DeepSeek V4 Flash284B params | 8 tok/s |\n20 tok/s |\n| Gemma 426B params | 140 tok/s |\n500+ tok/s |\n| FLUX.2 Klein1024² | Model · 4B~0.49 img/s |\nModel · 9B~0.92 img/s |\n| Z-Image-Turbo1024² | Model · 6B~0.22 img/s |\nModel · 6B~1.19 img/s |\n| Wan 2.2720p | Model · TI2V-5B~6.4 gen fps |\nModel · I2V-A14B~3.0 gen fps |\n| MiniMax H30.5 MP · native audio | Model · 33B~0.7 gen fps |\nModel · 33B~4.1 gen fps |\n| Cyberpunk 20774K · Ultra preset · DLSS 4 | 80 fps |\n110 fps |\n| Black Myth: Wukong4K · Cinematic · DLSS 4 | 60 fps |\n90 fps |\n| CapCut4K export | 75 fps |\n90 fps |\n| Premiere Pro4K H.264 export | 70 fps |\n90 fps |\n| Audio MemoryTTFA | ≤1.2 s |\n≤0.8 s |\n| BlenderCycles render | 2,900 samples/min |\n6,000 samples/min |\n\n†Open-weight models, served behind OpenAI- and Ollama-compatible endpoints.\n\nMeasured over Wi-Fi 7 on a production WiCi One — close to cloud rates of [20-40 tok/s](https://openrouter.ai/moonshotai/kimi-k3:nitro#providers).\n\nHow do we serve T-param LLM? [Learn more.](article/ssd-is-the-new-vram)\n\n†Open image models, served through the WiCi SDK's multimodal pipeline.\n\nMeasured over Wi-Fi 7 on a production WiCi One.\n\n†Games run unmodified on your own machine — not remote play: WiCi One attaches as a standard virtual GPU through the WiCi driver.\n\nMeasured over Wi-Fi 7 on a production WiCi One.\n\n†CapCut and Premiere Pro run unmodified in Driver mode — GPU effects and encoders offload over Wi-Fi. Audio memory runs on the WiCi SDK's local audio pipeline.\n\nMeasured over Wi-Fi 7 on a production WiCi One.\n\n†Blender runs unmodified in Driver mode, scored on its standard benchmark scene.\n\nMeasured over Wi-Fi 7 on a production WiCi One.\n\n-\n**API, SDK, and Driver modes.** Call it as an endpoint, build with the SDK, or attach it as a virtual GPU.[Build on WiCi](developers) -\n**Minimum overhead, strong performance.** The WiCi Protocol keeps wireless serving close to local speed.[Explore the technology](technology) -\n**Built on research.** A decade of wireless and AI-systems work, published at SIGCOMM, NSDI, and OSDI.[Read our research](research)\n\n## Tech specifications\n\n### Wireless AI Computing\n\n-\nCPU\nIntel Core Ultra 7 255H AirCompute GPU sharing\n-\nGPU\nNVIDIA RTX 5060 Ti · 16GB Configurable to NVIDIA RTX 5090 · 32GB Pricing for this configuration to be announced closer to launch\n-\nWi-Fi\nWi-Fi 7 · 4×4 MIMO · 320MHz Router-grade AirLink Multi-Link transport\n\n### Fully Flexible GPU Runtime\n\n-\nDriver mode\nWiCi Virtual GPU on macOS, Windows & Linux ZeroTrip caching & trace replay\n-\nSDK mode\nPython · JavaScript · Swift Out-of-the-box inference with multimodal support\n-\nAPI mode\nOpenAI-compatible endpoint Drop-in for existing clients\n\nWiCi One is fully application-transparent in Driver mode, inference-optimized through the SDK, and OpenAI-compatible in API mode. Whichever path you take, your applications get full computing flexibility.\n[Build on WiCi ](developers)\n\n### Unified Hierarchical Memory\n\n-\nStorage\nNVMe 5 · 4TB TurboStream weight paging\n-\nLLM\nTrillion-parameter models\n\nMeasured on WiCi One 5090 edition, close toKimi K3 2.8T 5 tok/s DeepSeek V4 Flash 284B 20 tok/s Gemma 4 26B 500+ tok/s [cloud rates of 20-40 tok/s](https://openrouter.ai/moonshotai/kimi-k3:nitro#providers)\n\nTurboStream turns 4TB of NVMe into working VRAM: trillion-parameter models, no cloud cost required. State-of-the-art research from our systems researchers and engineers, built into WiCi One.\n\n[Read more about our technologies →](technology)\n\nStarting at ~~$2,599~~ **$1,999** USDFor early signups only\n\n[Sign Up to Win WiCi One](#signup)\n\n##\nFast. Local.\n\n*Personal.*\n\nBuilt for low-latency interaction — voice, agents, and responsive AI experiences that feel like they're happening right next to you, because they are.\n\nAccess the best GPU over Wi-Fi instead of carrying it — every device stays cooler, lighter, and lasts longer.\n\nThe more personal AI becomes, the more sensitive the data it touches. Keep processing inside your own walls — local-first by default.\n\nHigh-frequency AI gets expensive when every interaction depends on cloud inference. Own the compute once — and share it across every device and everyone at home.\n\n##\nA wireless GPU runtime\n\n*for local AI apps.*\n\nWiCi One is a programmable endpoint on the wireless compute layer — run existing apps unmodified in Driver mode, build with the WiCi SDK, or call it through the API. At every altitude, your app keeps interaction, input, and UI on the device while GPU-bound inference offloads over Wi-Fi — model caching, streaming, and pipelining keep it real-time.\n\n-\n### Virtual GPU driver\n\nAttach WiCi One as a virtual GPU — existing applications run unmodified, fully transparent.\n\n-\n### LLM-optimized SDK\n\nBuild with full control — PyTorch and multimodal pipelines on wici.gpu, while I/O stays on the device.\n\n-\n### Drop-in API with trillion-parameter model support\n\nKeep your existing client — point it at WiCi One's OpenAI-compatible local endpoint.\n\n##\nRun AI where *life happens.*\n\n### Personal AI assistant\n\nPower fast, always-available AI agents across your personal devices — anywhere in the home or office.\n\n### Shared studio compute\n\nEquip a whole studio of lightweight laptops with one shared GPU for AI-assisted rendering, image generation, and video editing.\n\n### Voice and multimodal agents\n\nSupport real-time voice, audio, vision, and interactive AI experiences.\n\n### Smart space intelligence\n\nEnable speakers, cameras, robots, wearables, and AR glasses to tap nearby AI compute.\n\n### Developer prototyping\n\nBuild and test local AI applications using WiCi's protocol and SDK — from prototype to product.\n\n##\nA smart speaker,\n\n*free for early adopters.*\n\nA thank-you for our first customers — and a first glimpse of the local AI devices WiCi One can power.\n\n### Programmable.\n\nHackable hardware, real-time voice.\n\n### Local voice.\n\nPowered by nearby GPU, not the cloud.\n\n### App-paired.\n\nConfigure from the WiCi One App.\n\n### Open SDK.\n\nBuild local AI from day one.\n\n```\nIn summary\n\n                        The whole picture,at a glance.\n```\n\n### AI compute, closer to you.\n\n### Wi-Fi, GPU, SDK — in one box.\n\n### Real-time, by design.\n\n### Your data, your walls.\n\n### Run AI in every room.\n\n### A smart speaker, free for early adopters.\n\n## Stay in *the loop.*\n\nWe're building the wireless compute layer for personal AI — quiet by design.\n\nSign up for the newsletter — and **a chance to get WiCi One free** — plus a free smart speaker for early adopters.\n\n##\nQuestions?\n\n*Answers.*\n\n-\nNo. AI PCs put acceleration inside a single machine. WiCi One makes compute a shared resource all your devices access wirelessly.\n\n-\nNo — it complements cloud AI, taking on the workloads where local latency, privacy, and cost matter most.\n\n-\nIt is a bonus programmable device included for early adopters to experience local compute-enabled AI interaction.\n\n-\nDevelopers, researchers, AI builders, early adopters, and teams interested in personal local AI compute.\n\n-\nWe anticipate shipping starts in Q4. Availability will be announced to newsletter subscribers first — along with the details of the Pioneer plan and its top-10 free devices. Sign up to be first in line:", "url": "https://wpnews.pro/news/wici-one", "canonical_source": "https://wici.ai/wici-one", "published_at": "2026-09-03 02:35:18+00:00", "updated_at": "2026-09-03 02:51:52.886217+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-products", "ai-tools", "generative-ai"], "entities": ["WiCi One", "Ollama", "LM Studio", "Hugging Face", "PyTorch", "ComfyUI", "Open WebUI", "Cyberpunk 2077"], "alternates": {"html": "https://wpnews.pro/news/wici-one", "markdown": "https://wpnews.pro/news/wici-one.md", "text": "https://wpnews.pro/news/wici-one.txt", "jsonld": "https://wpnews.pro/news/wici-one.jsonld"}}