v0.1.11 — a model inside the browser, a finder for local ones, and multi-line WhatsApp messages

Two new ways to use ProofKey with no API key where the browser has no built-in model, such as Brave, and a fix for multi-line messages on WhatsApp Web.

A model inside the browser

Model in this browser (Qwen3.5 4B, no key) runs the model inside ProofKey itself, on your GPU through WebGPU (WebLLM 0.2.85 in a hidden extension page). Nothing is downloaded until you press its own button, which names the size and the host first: 2.4 GB from huggingface.co, pinned to one revision. After that, nothing you check leaves the computer. It needs WebGPU and about 4 GB of GPU memory.

It is offered live checking, Fix grammar, bullet points and Translate, the actions it measured well on. The rewrites are refused on it and say so. Measured in Brave on Windows with an AMD RX 6600: 13.0/14 on the live check over 10 runs with no false alarms, the same as Chrome's built-in model, and Fix grammar 11/12. One machine; details in MODELS.md.

This adds the offscreen permission (no install warning) and lets extension pages run WebAssembly. The model's compiled WebGPU library ships inside the package; its source repository states no licence, and the question is open at mlc-ai/binary-mlc-llm-libs#166. THIRD_PARTY_NOTICES.txt in the package says where each bundled piece comes from.

Find a model on this computer

Find a model on this computer, on the built-in model's card, looks for LM Studio, Ollama or llama.cpp on their usual ports (1234, 11434, 8080). It only looks when you click, and only on this computer. Then it adds the one it finds as your first provider. When Ollama refuses the extension, the card names the OLLAMA_ORIGINS setting that fixes it. Tested against a real llama-server; not yet against real LM Studio or Ollama.

Multi-line messages keep their line breaks

Correcting a WhatsApp Web message of several lines (typed with Shift+Enter) could join the lines into one or leave a stray word at the end. Edits are now checked against the whole field, and a multi-line reply keeps its line breaks in Lexical editors.

Also in this release

• When the default action has no shortcut, the notice suggests a free one and says where to set it.
• The shared language rule has an English example as well as a Spanish one. Local models had been answering English messages in Spanish on Summarize (Qwen3-4B: 2/4 kept English, now 4/4).

How it is known

On the released tree, npm run verify passed 497 checks and failed none. That covers unit checks, test:render and test:ext in Chromium. The in-browser model's test:ext checks run against a stand-in engine, because this machine has no GPU. The real WebLLM was loaded inside the built extension on a software WebGPU adapter. It was also run in Brave on Windows with a real GPU: the download reached Ready and Fix grammar worked on WhatsApp Web. That run was reported by the user, and its output text and latency were not recorded.