HesvorynChat.
Run a small language model in your browser. Stream the answer, stop it, edit the prompt and try again.
Open workspace →Think it through.
Keep it close.
Crypto research. Unfinished ideas. A model that can run right here.
No account. No API key for local models. You choose what leaves your device.
FOLLOW THE MESSAGE
Who can read
what you type?
Compare.
Run the same prompt with two local models. Keep the exact model name, response time and output side by side.
Try a comparison →Keep.
Export a readable transcript or encrypt a copy with a passphrase. Open it again when you want to continue.
Open the note vault →SMALL MODELS / REAL INFERENCE
Pick what
fits your device.
Models download only when you load one. Device memory and WebGPU support determine what can run.
Checking browser model support…
YOUR WORKSPACE
Start with
the actual question.
◈ Local · prompt stays in this browserNo model loaded
SAME QUESTION / TWO PERSPECTIVES
Compare the answer.
Keep the evidence.
Runs one model at a time to avoid holding two models in GPU memory.
MODEL A
Your first response will appear here.
MODEL B
Your second response will appear here.
TAKE YOUR NOTES WITH YOU
A copy.
Under your control.
Export the current conversation. Use a passphrase to encrypt the file, or save a readable transcript. No server vault or account is involved.
AES-GCM encryption with a PBKDF2-derived key. Keep your passphrase separately; the site cannot recover it.
⌗
Open a saved conversation.
WHAT THIS VERSION DOES
Clear boundaries.
Useful tools.
Local inference
The model runs through WebLLM using your browser’s WebGPU support. Model files come from the model publisher when you choose to load them. Prompts are not sent to a hosted inference server in Local mode.
Models this small can make factual and reasoning errors. Use them for drafts and explanations, then check cryptocurrency contract facts against primary sources.
Connected API
You can supply an OpenAI-compatible HTTPS endpoint, model ID and key. Requests travel directly from your browser to that provider. The provider can read the request; this mode does not claim a hardware enclave or zero retention.
Keys remain in memory and are excluded from exports. CORS restrictions, provider authentication and provider charges still apply.
Conversation storage
Conversations stay in this tab’s memory. Reloading clears an unexported conversation. Cached model files do not contain your chat. Encrypted exports contain a random salt, nonce and ciphertext.
The site does not perform wallet actions, approve tokens or promise investment outcomes.
Reproducible details
Every response shows the actual model and runtime mode. Comparison results include elapsed time. Generation failures remain errors; a scripted answer is never substituted.
Why does the first response need a download?
The model’s weights must be on your device before local inference can run. The page shows real download progress, and subsequent loads can use the browser cache.
Why can’t my browser load a model?
WebGPU, compatible graphics drivers and enough device memory are required. Try a current desktop browser or use your own API endpoint. There is no fake local-mode fallback.
Is the conversation anonymous or hardware sealed?
This version offers local inference and explicit external API access. It does not operate an anonymous routing network or an attested hardware enclave.