QuantumOrb holds every launch on Pump.fun and PumpSwap as a probability state over three outcomes: +60% first, −30% first, or neither within the hour. Six agents feed it signals that reinforce or cancel each other, the state updates with every trade, and it collapses into a trade only when the edge survives every cost. It works in the minutes after a launch, not in the first slot.
Real figures from Pump.fun, the largest memecoin launchpad on Solana. Under 1% of launches graduated in the spring; since BOOST became the default launch in late July 2026 it is a few percent a day. Either way the job is less about finding winners than about staying out of everything else.
Every star is a live Pump.fun launch; the points around it are its 20 largest holders outside the curve and the pool, read on Solana. A wallet that sits in the top 20 of several launches is drawn once and tied to each of them in red: one actor entangled with more than one launch, so what it does in one says something about the others. The creator, read from the bonding curve, is gold; wallets Judge has scored are green when they are in profit. Drag to explore, scroll to zoom out to the whole globe, click any star to see who is behind it.
Scout, Cartographer and Judge carry observations up into the Orb. Quant, Guard and Executor take its probability states down and collapse them into decisions; Guard also screens every launch on the way in. An hour after the Orb first scores a launch, the measured outcome goes back up as a training label, traded or not. Open a card to see how the agent works and what it is doing right now.
The Orb holds every launch it scores as a state over three outcomes: +60% first, −30% first, or neither within the hour. Each signal adds a signed push to that state: signals that agree reinforce each other, signals that conflict cancel. Move the signals, watch the state shift, then measure it and see the outcomes land in proportion to their probabilities. The weights are the core model's own, and Guard's lines and Quant's cost math are the same code the launch screen runs.
Every launch starts at the prior, about a 15% chance of +60% first, and evidence moves it from there. The arrows add head to tail, so you can see which signals reinforce and which cancel. TIME follows the speed of buying and −30% takes the rest, so the three always sum to one. Guard's veto lines are marked on the sliders.
Run it yourself: trace a trigger, a followed wallet's buy or strong order flow, through all seven steps as the state forms and collapses, pause it, open any line to see why, or type a run number.
When a followed wallet buys a coin, or its order flow turns strong, the agents answer one after another: what happened, who is behind the buyers, whether the buyer is worth following, what state the Orb holds, whether the trade pays after costs, whether anything is wrong, and how to land it. Most states collapse to a block or a skip; the ones that pass collapse into a trade entry. An hour later the launch is measured, traded or not, and its label goes back into training. No coin is named: each run carries only its number, and the same number always replays the same run.
The agents above do not guess. Each step is a model with a number you can check: how a probability state becomes a sized trade, how a wallet's belief state updates and decoheres, where a buy lands on the bonding curve, and whether the Orb's probabilities hold up when they are measured.
The Orb holds each launch as a state over three ways out of a trade: take-profit, stop or the time limit. Weighted by their payoffs they make the gross edge. Then come the costs: the launchpad fee in and out (1.25% a side on a Pump.fun curve: 0.95% protocol plus 0.30% creator; on a PumpSwap pool it starts at the same 1.25% and falls as market cap grows), the priority fee and Jito tip to land both the buy and the sell, and the price impact of our own buy and sell, computed exactly on the curve. The stop is priced where stops fill on a gap, about −36%. Only a remainder above +3%, a margin for the model's own error, with costs under 4% of the trade, gets a size: Quant searches 31 sizes for the one that makes the most SOL in expectation, at most a quarter-Kelly stake at the cautious end of Core's estimate and never more than 3% of the book, an average fill at most 1% above the current price, and the room Guard leaves in the book. On a Pump.fun curve the fees alone take 2.5% of a round trip, so in practice the 4% cost cap sets the size. A time-out is priced at its average exit, about −7%.
Judge does not count wins: memecoin traders often win a third of the time and live off the rare big run. It keeps a belief state for each wallet's profit per trade after fees: it starts from a typical wallet that loses a little and every closed trade updates it. Each trade counts at most +200% and each wallet's own spread is estimated, so one lucky token barely moves it; old evidence decoheres, losing half its weight every 21 days; and what copy-bots and our own one-second delay take is subtracted. A wallet is followed only when the chance its edge survives at our entry passes 90%.
A Pump.fun coin has no order book. Its price comes from a constant-product curve with 30 SOL and 1.073B tokens of virtual reserves: every SOL in raises the price for the next buyer, and at about 85 SOL the curve sells out, the coin bonds and moves to a PumpSwap pool. Before buying, Quant simulates the exact fill: the tokens its size gets, the average price over spot and how far it moves the price. Earlier is cheaper per token, but the same SOL moves a thinner curve more. The curve is the one fully deterministic part of the system: every probability upstream is priced through it exactly.
The Orb is a gradient-boosted model (LightGBM) that holds three chances: +60% first, −30% first or neither within 60 minutes. Every launch that passes Guard is measured: the barrier its price touches first in the hour after the model first scores it collapses the state into one label, whether it was traded or not. The hour comes from the data: in April 2026 coins that bonded organically took a median 9.8 minutes (bot-driven bonds, about a quarter, under one), so a coin still on the curve after an hour had lost its reason to hold. Launch rules keep changing, so the window is re-tested with every retrain. Labels retrain the model each night with a one-hour gap before the test days, its probabilities are checked on every scored setup, and a new version ships only if it beats the live model's recipe, trained on the same days, on days neither has seen, by more than the noise.
Chain indexer, wallet scores, launch screening, model decisions.
The code on GitHub and this site in public, with the model's training in the open.
$QORB launches and trades on the open market. The bot keeps training on a model book.
When the conditions on the Training tab are met, the bot gets real capital and trades from a public wallet.
Subscriptions paid in $QORB.
It never holds your funds or asks for your keys.
Judge reads the last 300 trades of every live Pump.fun launch, each with the wallet that signed it, and keeps every wallet with two trades or more. For each wallet and token it adds up the SOL paid on buys and the SOL received on sells. The result is the SOL it got back minus what it paid for the part it sold; a position counts as closed once it sold 95% of what it bought. Wallets that also sit in a launch's top 20 holders, or created one, are marked. Ranked by realized SOL.
| # | Wallet | Realized | Win rate | Closed | Buys / sells | Tokens | Median hold | Avg buy | Holds now | Last trade |
|---|
Pump.fun tokens from DexScreener's listings, refreshed every minute. A curve shows here once 5 SOL is in it and bonds to PumpSwap at about 85 SOL. Guard reads each token on chain: its authorities, its Token-2022 extensions, the 20 largest holders and the bonding curve, which gives the SOL in it and the creator. One red flag is a veto. A clean token's hour starts ten minutes after its curve passed 5 SOL, and the first barrier its price touches in that hour, +60%, −30% or neither, is its label, read from 1-minute candles.
| Token | Age | Stage | Vol · 5m | Liquidity | Mcap | Buys / sells · 1h | Price · 1h | Top 10 | Guard | Label · 60 min |
|---|
Every launch that passes Guard becomes a lesson: from the moment the model first scores it (a followed buy, strong order flow, or ten minutes after the Guard check if nothing triggers it), whichever of +60%, −30% or 60 minutes its price touches first is the measurement that becomes its label. Each night a challenger and the live model's recipe are trained on the same 14 days and tested on the 3 days after them, which neither has seen; the challenger has to win by more than the noise. This tab shows that process, and what still has to happen before it trades real money. It does not show what any trade made.
The bot stays in research mode until every condition below is met. They test the model where it trades, the model results and the risk limits, plus one step for the project itself: its code in public. Model results count only as met or not yet; no amounts are published.
Each dot is a launch, with no name, held as a state over three outcomes from the moment the model first scores it. Its price runs until it touches +60%, −30% or the 60-minute wall, and that first touch, the measurement, is its label. Traded or not, every launch counts, so the model also learns from the ones it skipped.
Skill is how much better the model predicts than a guess that knows only the average outcome (log-loss on 3 days that neither model saw). To keep the race fair, the live model's recipe is retrained on the same 14 days as the challenger, so the live model never tests on days it learned from. A challenger replaces it only when it is better by more than two standard errors; most nights it is not. One that passes is refit on the latest 14 days before it goes live.
All setups the model scored in a week are grouped by the chance it gave. Each dot is how often that group really hit +60% first; on the diagonal, measured frequencies match the stated chances. It is checked on every scored setup, traded or not.
Every closed trade updates a wallet's belief state, and older trades decohere, losing half their weight every 21 days. Wallets are added when the chance that their edge is real reaches 0.90 and dropped when it fades, when copy-bots take most of it or when they turn out to be one entangled group. Counts only; no addresses.
Every launch Guard blocked is followed for 24 hours anyway. Dead means down 90% or more from the price at the check. Blocked launches should die far more often than the ones it let through, and a few blocked ones still run: that is the price of caution.
Last updated: 11 October 2026. QuantumOrb is an experimental project, $QORB is a memecoin, and you can lose everything you put in. By using this site you accept these terms.
QuantumOrb is an independent, experimental software project. The site shows how an automated model studies memecoin launches on Solana. It is provided for information and entertainment only. It is not a company, a fund, a broker, an exchange or an investment service.
References to quantum concepts on this site, such as probability states, interference and measurement, describe how the model is framed.
Nothing on this site, in the Telegram bot or on any QuantumOrb channel is financial, investment, legal or tax advice, a recommendation, or an offer or solicitation to buy or sell anything. Model outputs are automated estimates and can be wrong. Do your own research. Any decision you make is yours alone.
If signals are offered, they are the same automated messages for everyone and take no one's personal situation into account. They can be late, wrong or missing. Acting on a signal is entirely your decision and your risk.
The bot's own trading and the signals are kept apart: the bot trades other tokens from its own wallet and never buys or sells a token it sends as a signal, so subscribers are never used as exit liquidity for its positions.
Payments for a subscription are final and cannot be returned.
QuantumOrb is not affiliated with or endorsed by Pump.fun, DexScreener, GeckoTerminal, Jupiter, Telegram, X, GitHub or any other platform, wallet or service it reads from or links to. Their names belong to their owners. QuantumOrb does not control them and is not responsible for their services, fees, failures or content.
Everything is provided "as is", with no warranty of any kind. To the fullest extent permitted by law, no one involved in QuantumOrb is liable for any loss or damage, including the loss of funds, arising from the site, the bot, the signals or $QORB.
The site, the bot, the signals or the project as a whole can be changed, paused or stopped at any time, without notice.
You alone are responsible for your actions and decisions, and you alone bear every loss that results from them, including the loss of all the funds you put in.
If your actions, or your breaking of these terms or of any law, lead to a claim against QuantumOrb or anyone involved in it, you, the user, pay for that claim and for any loss it causes. QuantumOrb does not.
You must be at least 18. Do not use QuantumOrb or buy $QORB where it is prohibited, including in the United States and in sanctioned countries. You are responsible for knowing and following your local laws and for any taxes.
This site does not use cookies and does not collect personal data. Its fonts load from Google Fonts, which, like any web server, sees your IP address. The Telegram bot, if you use it, stores only your Telegram user ID, your subscription status and the transaction you paid with, which is public on chain anyway.
QuantumOrb never holds your funds, never asks you to connect a wallet or sign a transaction on this site, and never asks for your private keys or seed phrase. It will never message you first. The only payment it ever takes is a subscription you start yourself in the official Telegram bot. Anyone else who asks you to send funds or tokens in QuantumOrb's name is a scammer.
Everything QuantumOrb does, in the order it happens: where the data comes from, how six agents turn it into a decision, how a trade is sized and closed, how the system learns, what you receive and what the token is for.
A quantum-inspired intelligence network for memecoins on Solana. It reads every launch, curve trade and pool swap on Pump.fun and PumpSwap, holds each launch as a probability state, scores the wallets whose results hold up, screens each launch for manipulation, and collapses the state into a trade on its own only when it still pays after every cost.
Copying a trader is one input among several, never the whole decision. QuantumOrb does not race snipers for the first slot; it works in the minutes after a launch, when the chain shows who is really buying.
Quantum describes how the model is built, not the hardware it runs on. Five ideas organise it, and each one is a concrete part of the system:
Seven steps, six agents and the Orb, in the order the work moves. The target is about one second from the trigger, a followed wallet's buy or a burst of organic buying, to a decision: roughly three Solana slots.
Reads every confirmed slot, about one every 400 ms, and turns each launchpad event into one clean record: token, wallet, size, price, slot.
Traces who funded each buyer, who landed in the launch slot or the same Jito bundle, and which wallets belong to the creator. It answers one question: how many of these buyers are really one actor?
Scores the wallets trading the coin on profit per trade after fees and copy-bots, not on win rate.
A gradient-boosted model (LightGBM) adds the agents' signals into one state of three chances: +60% first, −30% first, or neither within 60 minutes.
Simulates the exact fill and takes every cost off. Only an expected value above +3%, with costs under 4%, gets a size.
Screens every launch once 5 SOL is in its curve, so the core only scores clean ones, then checks again right before the order: new red flags and room in the book. A veto collapses the state to no trade; Guard can stop a trade but never start one.
Turns the collapsed state into an order: lands it through Jito and sets the three exits. Today it trades in research mode.
You can run example traces of this path in the Life of a decision terminal on the home page.
Pump.fun coins are minted with fixed supply and the creator cannot pull the pool's liquidity after migration, so Guard looks where manipulation actually happens:
On this site Guard's basic checks already run on chain for every live launch: mint and freeze authority, risky Token-2022 extensions, the top 10 wallets and the largest wallet outside the curve, and the creator's share, read from the bonding curve. One red flag is enough to block, and the reason for every block is written to the decision log. Guard checks again right before any trade, for new red flags and for room in the book.
Memecoin traders often win a third of the time and live off the rare big run, so win rate says little. Judge keeps a belief state for each wallet's profit per trade after fees: a Bayesian model that starts from a typical wallet that loses a little, fitted to all wallets with 20+ trades. Each trade counts at most +200% and the spread of each wallet's results is its own, so one lucky token barely moves it; a long record does.
Old evidence decoheres: a trade loses half its weight every 21 days, so a wallet that stopped working loses its rating instead of living off old wins. The best wallets are copied by bots within a slot or two, and QuantumOrb itself enters about a second later, so what copy-bots and that delay take is subtracted: the only edge that counts is the one left at our entry. A wallet is followed only when the chance that its edge survives all of that is at least 90%.
The Orb builds the setup's state mostly from its order flow: new buyers a minute, buys over sells, the share of fresh wallets and whether the top holders are concentrating. Each signal adds a signed contribution, so signals that agree reinforce each other and signals that conflict cancel, leaving the state near the prior. A followed wallet's buy is one feature among these, not the only way in: strong organic flow (8+ new buyers a minute, buys at least 1.4× sells, at most 40% fresh wallets) triggers a decision too. Bundle share in slot 0, the creator's record and a linked Telegram channel complete the picture. Since late July 2026 a freshly migrated Pump.fun coin gets five minutes of automatic BOOST buying, burned as it is bought; that demand is mechanical, so it is not read as buyers' interest. Its probabilities are checked against outcomes and recalibrated, so a 30% call should keep meaning about 30%; the calibration chart on the Training tab shows where it has been too sure.
Quant then weighs the three ways out of the trade, pricing the stop where stops really fill on a gap (about −36%, not −30%) and a time-out at its average exit, about −7% today and re-estimated from the labels, and subtracts every cost: the launchpad fee each way (1.25% on a Pump.fun curve as of September 2026), the priority fee and Jito tip on the way in and out, and the price impact of its own buy and sell, computed on the exact constant-product curve for the planned size. Fees fall with a pool's fee tier, and the priority fee and tip follow the network's load instead of a fixed buffer. A trade needs the expected value after all of it to clear +3%, and the costs to stay under 4% of the trade. Quant then picks the size that makes the most SOL in expectation, at most a quarter-Kelly stake computed at the cautious end of Core's estimate, at most 3% of the book, an average fill at most 1% above the current price, and the room left in the book.
Whichever barrier the price touches first closes the trade. Memecoins fall together when SOL does, so a single coin is never sized alone: the whole book is capped, and no new trades open for the rest of the day once the day's closed losses reach 5%. The day stop only blocks new entries; open positions keep their own stops. So the worst day is the 5% already lost plus every open position stopping out on a gap at about −36%: in theory, with the full 15% in play, about −10% plus costs; at today's sizes, with a few percent in play, nearer −6.5%.
In practice the size is set by the 4% cost cap. On a Pump.fun curve the launchpad fee alone takes 2.5% of a round trip, so only about 1.5% is left for the priority fee and Jito tip in and out and for the price impact of our own buy and sell. That allows roughly 0.2 to 0.7 SOL a trade, and when the network is busy no size fits until about 20 SOL is in the curve. Six open positions use only a few percent of the book. Profit is capped by costs and by how much the curves can absorb, not by the size of the book: more capital does not mean proportionally more profit.
Every launch that passes Guard is measured: from the moment the model first scores it (a followed buy, strong order flow, or ten minutes after the Guard check if nothing triggers it), the first barrier its price touches within the hour, take-profit, stop or time limit, collapses its state into a label, whether it was traded or not. Labelling only the trades it took would teach the model nothing about the setups it turned down. The hour starts at that first score, not at the 5 SOL check: on a Pump.fun curve, 30 SOL of virtual reserves floor the price about 27% below its level at 5 SOL, so a −30% stop could never be reached from there. Each launch counts once, from its first score: a later trigger on the same launch adds no new label, so busy launches do not outweigh quiet ones. The labels retrain the model every day on a rolling 14-day window, so it keeps up with a market whose rules and fees change often.
A one-hour gap separates the training days from the test days, so no label straddles both. Each night the challenger and the live model's recipe are trained on the same 14 days and tested on the next 3 days, which neither has seen, so the live model never tests on days it learned from. The challenger replaces it only if it is better by more than twice the day-to-day noise. A model that passes is then refit on the latest 14 days before it goes live, so it never misses the newest labels. Its probabilities are checked on every scored setup, not just the ones it traded, so a 30% call keeps meaning 30%. The Training tab shows every night's run and the calibration.
Once the bot trades live, signals are sold as a subscription in a Telegram bot. Each one carries:
Until then no signals are published, on this site or anywhere else. The Training tab shows how the model learns instead: the labels it collects, every nightly model run, shipped or rejected, and whether its stated chances match what happened. The bot's own trading and the signals are kept apart: the bot trades other tokens from its own wallet and never buys or sells a token it sends as a signal, so subscribers are never used as exit liquidity for its positions. QuantumOrb never holds your funds and never asks for your keys; what you do with a signal is your decision.
QuantumOrb covers token discovery, wallet research, risk screening and model-based decisions. The Tokens and Traders tabs, event stream and wallet graph use data from DexScreener, Solana and GeckoTerminal. The Training tab displays model diagnostics and readiness milestones.
Indexer, wallet scores, launch screening and model decisions.
The code on GitHub, then this site in public with the model's training in the open.
$QORB launches and trades on the open market. The bot keeps training on a model book; how it learns is public.
When the conditions on the Training tab are met, the bot gets real capital and a public wallet.
Subscriptions paid in $QORB.
$QORB is the project's token. Its planned use is one thing: paying for a subscription to QuantumOrb's signals once they are sold.
No. Today QuantumOrb trades only in research mode and publishes no signals. Once it trades live, signals go to subscribers in a Telegram bot, and what you do with one is up to you. The live wallet on the roadmap is QuantumOrb's own, public wallet; it never touches yours.
The Training tab shows how the model is doing without listing any trade: each night's run against the live model, shipped or rejected, and whether its stated chances match what happened. Once the bot trades live, its wallet's transactions are on chain for anyone to verify.
Pump.fun, and PumpSwap once a curve completes.
The first slot is a race of infrastructure and bundles, and most of what happens there is the launch group itself. QuantumOrb works in the minutes after, when the chain shows who is really buying.
A Telegram account and $QORB to pay for the subscription, once subscriptions open. They open only after the bot moves from research mode to live trading. Until then no signals are published anywhere; the Training tab shows how the model learns.
No. The token is for paying for signals. It gives no share of profits, no dividends and no claim on the bot or its wallet, and the bot's profit is never used to buy it back. Its price can fall to zero.