Charge is an open 30B model that reads Pons launches on Robinhood Chain. It learned from launches and what happened to each one next. On launches it had never seen, it beat Claude Opus 5.5 at the question trenchers ask all day: will this double before it halves? In a fraction of the time (3.6× faster) and at 40× lower cost per read.
Charge builds the facts card from the public tape and reads it, live. ETH-paired Pons launches up to 3 days old.
A few reads a minute per person. Not advice.
These are Charge's real reads of launches it never saw in training (October 10–12). We picked them to show every kind of outcome, misses included. The full numbers are further down.
Same facts, same question: will this launch double before it halves in the next 24 hours? Scored on launches none of them had seen.
How we tested. snapshots of live Pons launches from October 10–12, none seen in training; every launch that doubled first is included. Each model saw the same facts. The Claude models also got a primer on how Pons works and the 6% base rate, and rated 20 snapshots at a time. AUC weights each launch once. Test numbers come from Charge 30B trained only on launches before October 10; the released model is the same recipe trained on every launch through October 12.
Charge ranks a launch against every live launch it learned from: 90% means busier and stronger than 90% of them. Here is what launches in each band did over the next 24 hours, from Charge's own reads on the test days.
of ⚡ reads halved before they doubled; doubled first. The typical one dropped to of the read price within a day. Charged is a reason to watch, not to chase.
Test days: October 10–12, 2026, one launchpad. Small samples, and markets change.
Public trades of every ETH-paired Pons launch since October 5 that got at least one real trade (about a quarter never do): price, flows, buyers, sellers, holders, the dev's own trades, ticker copies, socials. Nothing private.
Each launch read at fixed ages and every 10 minutes while it traded: 65,542 moments, each labelled with what happened in the next 24 hours.
A gradient-boosted model learns the odds from those labels, with every launch weighted once so a long-running coin can't dominate.
Every answer is written from the data by templates: the band, the measured outcome mix and the facts that moved the odds for launches like it. No text from another LLM.
Qwen3-30B-A3B fine-tuned with LoRA (attention and experts) on examples, including general chat turns so it still talks like a model, then merged into open weights.
Trained on launches before October 10, tested after. Scored against what really happened, and against Claude Opus 5.5 and Sonnet 5.5.
Every Charge model is built the same way: public tape, real outcomes, an exam on launches it never saw, and a head-to-head with frontier models. Here's what's next.
Launches hours to weeks old that start trading again. Is the wakeup real, or one wallet pushing a dead chart?
Robinhood Chain · Pons · next
Coins launched straight into Uniswap pools on Robinhood Chain, where most of the action outside Pons happens.
Robinhood Chain · direct launches · next
The biggest trench. The same read for pump.fun launches, trained on hundreds of thousands of them.
Solana · pump.fun · after
No. Charge reads public trading facts and tells you how launches like this one did next. Even its hottest band halves first more often than it doubles. What you do with a read is your call.
No. It reads the facts you give it: age, market cap and ATH, trades, buyers and sellers per window, holders, the dev's own trades, first-30-second buyers, the creator's other launches, socials. The Charge bot fills those in from a contract address.
Because the hottest tapes are the most crowded. Busy launches move in both directions: the runners come from the top bands, and so do most of the round-trips. A high charge says "this is where the action is", not "this is safe".
It doesn't read contracts, so honeypots and hook traps need a sell simulation. It can't see the team, the news or who is about to shill. It was tested on two days of one launchpad. It will be wrong, often.
$CHARGE launched on the Surplus LLM launchpad. Its creator fees go to a fixed splitter: 60% to the hosting vault that keeps Charge live on the Surplus API, 30% to the builder, 10% buys and burns $SURPLUS.
The Surplus API serves the exact commit $CHARGE launched with. Newer versions trained on fresh tape, if we make them, will be published on Hugging Face.
The launch wallet can't buy on Surplus launches. Any dev buy is made from a public wallet, listed here.