Every LLM demo I have ever seen has the same fine print: bring your own API key, or the demo is rate-limited, or your prompts go to our server. So I built the opposite demo.
DevKit's AI Playground is a chat with a real language model that runs 100% in your browser. No backend, no API key, no cost per token, and — because there is no server at all — nothing you type can leave the page. It's a static HTML file on GitHub Pages, like the rest of DevKit.
How it works
- Transformers.js v3 does the inference, inside a Web Worker so generation never freezes the UI.
- The model is SmolLM-135M-Instruct, 4-bit quantized (
q4f16). The weights file is ~117 MB; with the tokenizer the one-time download is ~121 MB, and your browser caches it for future visits. - It tries WebGPU first (your graphics chip) and falls back to WASM on the CPU.
- Tokens stream in as they're generated, with a live tokens/sec readout — which is either a fun toy or a benchmark, depending on how serious your hardware is.
Click "Load the model", wait for the one-time download, and chat. That's the whole signup flow.
Honest limits
It's a 135M-parameter model that lives in your RAM. It writes plausible text, answers simple questions, does little creative tasks — and cheerfully makes things up. It is not GPT-class, and the page says so up front. The point isn't to replace an API; it's to show how far "the browser is the runtime" has come, and to give the LLM-curious a playground with literally zero setup and zero bill.
DevKit's other LLM tools are the grown-up counterparts: a cost calculator that prices 14 real API models with cached-read/cache-write splits, and a tokens/sec simulator. The playground is what those calculators look like when the model is small enough to be free.
Also new in this release
Four more client-side tools joined the kit: a JWT decoder (decode-only, and it says so — the signature is not verified), a cron expression explainer (plain English + next 5 run times), a color palette generator, and a Pomodoro timer whose countdown runs in the tab title. Same rules as everything on DevKit: no uploads, no accounts, no tracking, and every calculator's math is checked by automated tests (310 checks and counting).
Try it and tell me the tokens/sec number your machine gets — and which model you'd want to see in the playground next. SmolLM2-360M? Qwen? The download budget is the boss fight.
Top comments (0)