DiffuseChat

DiffuseChat

An on-device diffusion LM (LLaDA-8B, ~5.3 GB): the reply is not written left-to-right — a masked canvas denoises in place, tokens popping in in parallel, lowest-entropy first. KitDiffusionLM streams a snapshot per step (still-masked positions as ░) so you watch it happen.

The whole ML surface is two calls:

let dlm = try await KitDiffusionLM(catalog: "llada-8b")
let reply = try await dlm.reply(to: "What is the capital of France?") { step in
    print(step.text)   // the live canvas, ░ = still masked
}

Run

xcodegen generate
open DiffuseChat.xcodeproj

Run on a Mac (the shipped bundle is macOS-only today). First launch downloads the model from the Hugging Face Hub; later launches load from cache.

CLI (macOS, no Xcode)

swift run diffuse-cli --prompt "What is the capital of France?"

The live canvas streams to stderr; the final reply is stdout (--quiet for reply only).