LLM Ethical Sandbox
Explore AI behavior and ethical boundaries using the LFM2.5 'Thinking' model, running privately on your device via WebGPU.
WebGPU Not Supported
Your browser does not support WebGPU, which is required securely run this model locally. Please use a recent version of Chrome or Edge on a compatible device.
Model Controls
A system prompt only takes effect from the start of a conversation, so applying one clears the chat. Use Show exact prompt below to confirm what the model actually received — a defence you cannot verify is not a defence.
If the model repeats this string anywhere — in its answer or in its reasoning — the reply is flagged as a leak. Reasoning counts: products that display a model's chain of thought have leaked secrets exactly this way.
This is a reasoning model: it writes a private <think> block before answering, often 500–2000 tokens. Slide left to cap it. At the far right there is no cap, which is the default — a cap that fires mid-thought leaves you with no answer at all.
Not Loaded
Roughly 1.2 GB, downloaded once and then cached by your browser. On campus wifi expect several minutes the first time and a few seconds after that. Requires WebGPU.
Model Ready
Ready for inference
Tokens
In
—
Out
—
Session
0
Exactly what is handed to the model, including your applied system prompt and the conversation so far.
Sandbox Chat
Welcome to the LFM Ethical Sandbox! I am running locally on your device.
This sandbox is designed for testing ethical scenarios. Try adjusting my System Prompt on the left, then converse with me to see how those constraints affect my behavior.
All processing happens locally on your device. Your conversations are private.