I previously got this wrong because I relied on ChatGPT (or maybe it was Perplexity?), but Edge actually supports only about 4k tokens in the input context, while Chrome allows around 9k tokens — nearly double. The only real advantage Edge still has is that it’s freely available, meaning it can also be used in tools like Ollama or Jan AI. You can check the model’s true quota like this — the output shows how much of your input limit has already been used: const session = await LanguageModel.create(); console.log(`${session.inputUsage}/${session.inputQuota}`);

Today I was able to run a nice comparison between the Prompt API in Edge (SLM Phi-4 Mini) and in Chrome (SLM Gemini Nano) — same system prompt, same settings: - Edge makes noticeably more grammar, expression, and word choice mistakes than Chrome, at least in German. - Edge doesn’t follow rules very well, even when they’re explicitly defined in the system prompt — and even less so when they appear as initial chat prompts. It tends to rely more on its own training data. - Chrome, on the other hand, adheres closely to instructions and produces far fewer language errors in German. - Chrome is faster to initialize — the session is created more quickly. - Edge can hold more tokens in context, but that doesn’t help much if it largely ignores the information. This could be an advantage when the context contains lots of new data unfamiliar to the model, but in my use cases (product advisor systems), it didn’t make a difference. Overall, Gemini Nano currently feels much more mature and useful than Phi-4 Mini. That might change later — both are still experimental — but for now, only Gemini Nano is truly practical.

If the MCP command throws an error like “permission denied” — for example, when running the Chrome DevTools MCP server with npx -y chrome-devtools-mcp@latest — it might help to clear the npx cache (that’s what worked for me): rm -rf /Users/userName/.npm/_npx && npx clear-npx-cache In my case, it worked right away afterwards. How did I figure it out? I asked the AI (Cursor) — and in agent mode, Cursor even executed it automatically.

By the way, if you’ve landed here and don’t yet have a solid foundation in web development, feel free to check out my WebDev Community. There you’ll find many free courses and resources to help you take your first steps (and a bit further): Learn WebDev

Welcome! This category is for everyone looking for English resources on AI development. Here you’ll find: - Free tutorials (Prompt API, WebLLM, local models) - Starter projects with code examples - Links to English courses and updates Have a look at the classroom, I will be adding courses and more resources there step by step. If you have feedback or requests, feel free to comment below 🙌