Noviq AI

The First Private Inference network powered by contributed compute

0x61894EcfceE12f1842A41Fcb14599Ed340141C5b

[Noviq AI @NoviqAIxyz]

Introducing NoviQ AI - the first private inference network powered by contributed compute.

Instead of renting GPUs from centralised providers, anyone can contribute compute to power AI inference while preserving user privacy.

[Noviq AI @NoviqAIxyz]

Private inference, actually proven on @RobinhoodApp

https://t.co/jywW38NgHi

[Noviq AI @NoviqAIxyz]

Docs are now live at :

https://t.co/Gx1r8weJHn

[Noviq AI @NoviqAIxyz]

Welcome to Noviq

[Whale Insider @WhaleInsider]

JUST IN: Noviq AI just proved private inference works over a decentralized GPU swarm. No node sees your prompt. Contribute a GPU, earn real yield. Live on Robinhood.

[Whale Insider @WhaleInsider]

JUST IN: Noviq AI just proved private inference works over a decentralized GPU swarm. No node sees your prompt. Contribute a GPU, earn real yield. Live on Robinhood.

[Noviq AI @NoviqAIxyz]

Private inference. We measured it. 🔐

On a real open-weight LLM (Qwen2.5-0.5B), a node holding just one secret share of the computation recovers 0% of your prompt - even with the public weights and the exact attack that breaks naive obfuscation - while the model still runs correctly (~97% token agreement).

CPU-runnable in seconds. Code + receipts 👇

https://t.co/xDBmM4zHe7

[Noviq AI @NoviqAIxyz]

@King_Memento https://t.co/NuU7K4PS2c

[Noviq AI @NoviqAIxyz]

Hi everyone. We want to address what happened today.

Shortly after launch, we noticed a single wallet had accumulated a very large portion of the token supply. I became concerned that allowing trading to continue under those conditions could put the rest of the community at significant risk.

Based on what we saw, we decided to delay the next steps rather than push forward immediately. I know this caused frustration, and I’m sincerely sorry for the confusion and disappointment.

My priority has always been protecting the project and the community. Over the coming hours, we’ll share a full breakdown of what happened, what we’ve learned, and the plan going forward so everyone has complete transparency.

Thank you to everyone who has been patient and supportive. I’ll continue to communicate openly as we work through this together.

[Noviq AI @NoviqAIxyz]

Introducing NoviQ AI - the first private inference network powered by contributed compute.

Instead of renting GPUs from centralised providers, anyone can contribute compute to power AI inference while preserving user privacy.

[Noviq AI @NoviqAIxyz]

Start here:

Website: https://t.co/jywW38MIRK
Documentation: https://t.co/tuOfq8ahs7

[Noviq AI @NoviqAIxyz]

0x61894EcfceE12f1842A41Fcb14599Ed340141C5b

[Noviq AI @NoviqAIxyz]

Hi everyone. We want to address what happened today.

Shortly after launch, we noticed a single wallet had accumulated a very large portion of the token supply. I became concerned that allowing trading to continue under those conditions could put the rest of the community at significant risk.

Based on what we saw, we decided to delay the next steps rather than push forward immediately. I know this caused frustration, and I’m sincerely sorry for the confusion and disappointment.

My priority has always been protecting the project and the community. Over the coming hours, we’ll share a full breakdown of what happened, what we’ve learned, and the plan going forward so everyone has complete transparency.

Thank you to everyone who has been patient and supportive. I’ll continue to communicate openly as we work through this together.

[Noviq AI @NoviqAIxyz]

Most people picture "decentralized AI" as some abstract diagram. This is what it actually is.

A single chip. A board. A shard. A full rack of contributed GPUs, thousands of machines, none of them in a datacenter Noviq owns. Every piece of hardware in this picture belongs to someone who plugged it in and started earning.

Here's what that network is actually built to do differently. Every decentralized inference project says "uncensored, decentralized, private." Two of those are provably true the moment GPUs start contributing. The third one, privacy, has never been demonstrated by anyone in this space. Sharded inference splits a large model across contributor GPUs so no single machine holds the whole thing. That solved scale. It didn't solve what happens to your prompt as it crosses machines you don't control.

That's the part Noviq is built around. Not claiming privacy and moving on, but measuring exactly what a node can see, hardening the points where it matters most, and only calling something private once there's a number behind it.

Compute. Distribute. Infer.
Privacy you can verify, not just promise.

[Noviq AI @NoviqAIxyz]

Noviq's privacy claim is measured, not asserted.

A reconstruction adversary attempts to recover the original prompt from the activations an untrusted node actually receives. Defenses are evaluated against that adversary directly, and the recovery rate has to fall below a declared threshold before a run counts as private. Each run produces a privacy receipt recording that number.

Inference itself is sharded across contributed GPUs communicating over the open internet. No single node holds the full model, and no single operator holds the network.

There's no policy layer between a request and the model. Billing runs through a token rather than an account, so there's no persistent prompt history to retain or query later.

Every run also produces a correctness and privacy receipt a third party can independently check: which nodes served the request, measured WAN latency between them, a hash of the output, and the worst-case token recovery across the untrusted nodes involved.

[Noviq AI @NoviqAIxyz]

We wanted to prove PRIVATE inference on open-weight LLMs - the machines running your prompt shouldn't be able to read it.

Attempt 1: secretly rotate the activations. Then we broke our own method:
on open weights, a node holding the public weights W and their rotated copy W·R just solves R = W⁻¹·W' in closed form. Prompt recovery → ~100%. 💀

Attempt 2: secure multi-party computation. Each activation is split into two random shares held by non-colluding nodes. One share = uniform noise = ZERO information. Linear layers with public weights stay free.

Results (chart):
→ prompt recovery from one share: 0.2% (= pure chance)
→ the same attack that killed v1, replayed: still 0.2% (no secret to recover)
→ output correctness: 99.6% token match vs plaintext
→ hand an attacker BOTH shares (collusion): 99.2% — privacy = non-collusion

Per-share privacy here is information-theoretic, not "hard to crack."

Honest scope: toy model, honest-but-curious + non-colluding; MPC replicates
compute and adds network rounds. Malicious security + WAN cost next.

Private inference — measured, not promised.

[Noviq AI @NoviqAIxyz]

https://t.co/h2SoysPA9c