$DECENTRAL is live on Solana.
949aaReg5QnEKbde41BBj5Z45uU2Y8yhB4bZaif4pump
Decentral is a decentralized AI inference network on Solana that connects idle GPUs with people who need private, censorship-free inference. Prompts are encrypted client-side, jobs run on independent providers, and every payout and settlement lands on-chain where anyone can audit it.
Learn more: https://t.co/s2T3pdiqmj
We are currently facing high traffic on our platform, your request may take longer than usual to resolve.
We are working on expanding our infrastructure right now. Sorry for the inconvenience!
We'd rather show than tell.
Contracts public. Architecture documented. Commits land in the open, tests and all. If something we say doesn't match what the code does, you can see it yourself.
That's the deal we want: verifiable.
https://t.co/WoEWXDTlmU
We just crossed 80 users!
Small number, but real people. Every one of you is early to something we think matters, and the questions and feedback so far have already shaped where this goes.
Thanks for being here at 80. Let's see where it goes from here.
What happens when you send a message on Decentral:
Credits lock in escrow on-chain. The job routes to an eligible worker. It runs the model and streams tokens back. At the end, one transaction pays the worker from that escrow.
No middleman holds the funds. The program does.
Decentral is one network with four ways in.
Chat, if you just want private inference.
Earn, if you have a GPU to put to work.
API, if you're a developer who wants a drop-in endpoint.
Explorer, if you want to audit it all on-chain.
Same compute layer underneath. Pick your door.
Right now millions of GPUs sit idle while people are told AI compute is scarce and expensive.
Decentral closes that gap: connect the compute that exists to the people who need it, price it in the open, let anyone join either side.
The goal isn't a better AI company. It's making sure using AI never requires anyone's permission.
Open models on independent compute, settled in public. No gatekeeper deciding what you can ask, no single provider who can turn you off.
That's the whole point of Decentral.
We're going live on @Pumpfun in an hour!
Livestreaming the actual development, walking through the roadmap and where we're taking Decentral next, and answering whatever you want to ask.
Come see how a decentralized inference network gets built, in real time. See you there.
https://t.co/2fiX03M9fS
We're going live on @Pumpfun in an hour!
Livestreaming the actual development, walking through the roadmap and where we're taking Decentral next, and answering whatever you want to ask.
Come see how a decentralized inference network gets built, in real time. See you there.
https://t.co/2fiX03M9fS
Second test suite this week: the reward-distributor.
Checks what must not break: USDC lands in the vault on epoch open, a claim pays the right pro-rata share, a wrong epoch id rejects, a double claim fails.
Money math deserves tests that break it.
https://t.co/Qm8d7DtmhM
New commit: a full test suite for the job-router.
The tests that matter most are the ones that should fail: posting against an escrow you don't own, claiming a taken job, cancelling someone else's. Each must reject correctly.
The bad path is the work.
https://t.co/EQNKIfbDqc
The quiet advantage of open models: they keep getting better and you don't pay for the upgrade.
When a stronger open model ships, providers serve it and you use it. No vendor keeping you on last year's version until you pay more. The ceiling rises on its own.
Cheap inference for easy stuff, strong models for hard stuff, and you pick per request.
Most platforms push one expensive default. On Decentral the model menu is yours: send a quick question to a small model for a fraction of the cost, save big models for when they earn it.
Inference gets streamed, which makes billing interesting.
You don't know the final token count until the response finishes. So metering happens live as tokens flow, and a single settlement lands at the end against the escrow that was locked up front. One transaction per job, not per token.
A small implementation detail we care about: the reward math runs in u128.
When a staker claims their epoch share, the program multiplies their stake by the pool before dividing by total stake. Do that in u64 and a large pool overflows. So we widen to u128 for the multiply, divide, then narrow back, and reject anything that rounds to zero.