Real AI, right in your browser.Nothing to install. Nothing leaves the tab.
Free during the beta — no account, no sign-up.
Click a link — it runs in the tab, on your GPU.
Fully local. Your words stay in the browser.
Your chat saves — and follows you to memown.
Runs on your own GPU, offline — no cloud servers spun up to answer you.
Qwen2.5-7B (Q4_K_M GGUF) generating in a Chrome tab. First token in 0.95 s.
measured · 2026-06-10Qwen2.5-0.5B (Q4_K_M GGUF), same laptop, same browser.
measured · 2026-06-10While you chat, nothing goes out.
by designWorks after the first load, no connection.
measuredMeasured on 2026-06-10 on Windows 11 with an NVIDIA RTX 4070 Laptop GPU (8 GB), in Chrome (WebGPU): greedy decoding, 128 tokens, 18-token prompt, int8-activation decode path (the same trade-off as llama.cpp's default GPU kernels). Speed depends on your GPU. Method and raw results →
Every number is measured and public. Read the raw runs and reproduce them yourself.
A private AI chat that runs in a browser tab. The model runs on your own GPU through WebGPU: nothing to install, no account, and your conversations stay on your device. diciz is made by ZMLabs, in Sète, France.
It depends on your GPU. Measured on 2026-06-10 on a laptop with an NVIDIA RTX 4070 (8 GB), in Chrome: Qwen2.5-7B (Q4_K_M) generates 31.2 tokens per second, with the first token after 0.95 s on an 18-token prompt; Qwen2.5-0.5B generates 94.2 tokens per second. Integrated GPUs are slower. The raw results are public on GitHub.
Your prompts, chats and memory stay in the browser. The network is used to download the model weights once, from the URL you choose (for example Hugging Face), and a few libraries from jsDelivr: those hosts see a file download, never your conversations.
Yes, once the model has been downloaded: the weights are cached in your browser (OPFS), and diciz keeps working without a connection.
You paste the URL of a GGUF model. The public benchmark ran Qwen2.5 (0.5B, 1.5B, 7B), Llama-3.2-1B, Gemma-2-2B and Phi-3.5-mini; the model has to fit in your GPU's memory.
It is free during the beta, with no account, in the browser and for the desktop app download.
diciz is built by ZMLabs — a deep-engineering studio in Sète, South of France, making powerful software accessible.
Open a tab. It runs — and it remembers.
One account across memown, diciz and vrampilot.
Preview — accounts aren't open yet.
New here? Create an account