FAQ

Two layers. First, everything to dLLM travels over TLS (HTTPS), like any secure site. Second, for any model labeled End-to-end encrypted (our independent workers), your prompt gets an extra seal so the people running those workers can't read it:

• Each worker generates a fresh encryption keypair when it starts and publishes only its public key.
• The coordinator seals your prompt + conversation to that worker's public key before forwarding it, so it crosses to the worker as ciphertext that only that worker can open — no worker operator can read it off the wire or from logs.
• The worker decrypts it in memory only long enough to run the model; nothing about your prompt is stored.

Be precise about who sees what: no worker operator can read your prompts — and since workers are run by many anonymous people, that's the threat this closes. The model's reply is sealed too — the worker encrypts every token to your browser, so we relay an answer we can't read (your device decrypts it). The one thing we briefly see is your prompt — only for an instant, to screen the one hard line (no CSAM), route, and meter — then we keep nothing. We bill from the token counts the worker reports, not from reading your text. So today: no operator sees your prompt, we glimpse it only to screen and keep nothing, and we never see the response. The single remaining gap — our momentary view of the prompt — closes in the hardware-enclave (“confidential”) tier.

Your saved chat history is a separate, stronger guarantee: it's encrypted on your device with a key derived from your 12 words, and we store and sync only ciphertext — that history we genuinely cannot read.

Models labeled Anonymous (the big hosted models) are different: your prompt is relayed in plaintext to a hosted provider to answer. Never tied to your identity — no account, we proxy it — but not sealed. Every model shows its label, so you choose.

A popular model isn't tied to one machine. Anyone can run a worker for it, so a model is served by as many workers (replicas) as have joined.

• Routing: each request goes to a worker offering that model, picking the cheapest live offer, and fails over to the next automatically if one errors or drops — so a single slow or offline worker never breaks the model.
• Concurrency: each worker handles several jobs at once (its operator sets the cap), and the coordinator spreads load across the available workers and their free slots.
• Scaling: more demand means workers earn more, which draws more workers in — capacity for a model is the sum of everyone serving it. Encryption isn't a bottleneck, because each job is sealed to whichever worker takes it at request time.

Today each request is handled whole by one worker, which is great for any model that fits on a single GPU. Genuinely massive frontier models — too big for any one card — are a later tier where a model is split across a pool of workers (on the roadmap). For current models, scaling is simply more workers.

Your saved history: no — it's encrypted on your device with your 12-word key and we store only ciphertext. The model's reply on an end-to-end-encrypted model: no — it's sealed to your browser, so we relay it without reading it. Your prompt: no worker operator can read it (sealed to the worker), and we (the coordinator) see it only for an instant to screen the one hard line (CSAM), route, and meter — then keep nothing; we store no prompts or responses. Anonymous (hosted) models are proxied to a provider. Nothing — account, email, or otherwise — links any of it to your identity.

Be clear-eyed here: dLLM is a normal website, not a Tor service, so your IP reaches our server at the network level like it would for any site. We don't keep IP logs and nothing ties a request to your identity — but if you want network-level IP privacy, use Tor or a VPN. dLLM works fine over both.

Best rule: treat a prompt like something an anonymous stranger might read. The system protects who you are far more strongly than what you type.

• On an end-to-end-encrypted model your prompt is sealed so no worker operator can read it — but the worker decrypts it to compute the answer, so the person running that box could in principle see the text (they never learn who you are). We, the coordinator, also see it briefly to screen for CSAM, then keep nothing.
• On an anonymous (hosted) model it's the other tradeoff: a large provider processes the text rather than a random operator, and again it's never tied to your identity.

So skip details you wouldn't want an anonymous third party to possibly see — real names, addresses, account numbers. Hardware-enclave workers, which would hide the prompt even from the operator, are on the roadmap.

The balance is gone — there's no "forgot password," because there's no identity for us to reset. Your 12 words are the only key to your account and history, so write them down and keep a backup. That's the trade for being unidentifiable.

You buy prepaid credit and spend a small amount per use — no subscription. Top up by connecting a Solana wallet and paying in USDC (more crypto options are rolling out). Credit is closed-loop: it only ever becomes more dLLM credit, never cash back — think gift card, not bank account. You can also send credit to anyone as a one-time voucher.

Yes, on every topic but one: absolutely no sexual content involving minors. That's enforced and non-negotiable — such requests are blocked and never processed. Everything else, the models discuss openly.

Yes — dLLM speaks the OpenAI API, so tools and code that already talk to ChatGPT-style endpoints work unchanged. Your account key works as a full-access key; for app-specific keys you can cap and revoke, mint them on the API page.

Independent operators with GPUs — anyone, including you. A worker does decrypt your prompt to run it (it has to, to compute the answer), but it never learns who you are, and the prompt is sealed so only the assigned worker — not the network — ever sees it. Today this is trust-based; verification and staking are on the roadmap. Want to earn by running one? See the worker setup guide.

The one hard line is enforced upstream, not on your machine. Every prompt first passes dLLM's CSAM-prevention gate at the coordinator — a keyword screen plus an LLM classifier that escalates anything touching minors. Anything it flags is blocked and never dispatched, so it never reaches your worker.

• What does reach you arrives sealed — encrypted to your worker, decrypted only long enough for your local model to compute the answer, and the reply goes back the same way. You never learn who sent it, and nothing on your box ties a prompt to a person.
• Optional web search runs over Tor, so your home IP is never exposed when a job looks something up.
• Beyond that one hard line the network is intentionally uncensored, so a worker may serve edgy-but-legal prompts — you're providing compute, not curating content. You choose which models you run, and can stop anytime.