charge-ai/charge-rh-30b-v3

Charge v3 (30B)

Charge reads a Robinhood Chain token's public facts and chart and says how live it is (its charge), what the chart is doing in plain words, and how tokens like it did over the next 24 hours, from the price right now. One model for ETH-paired Pons launches and direct launches into Uniswap pools, at any age up to 10 days.

Site: https://chargeai.bid · X: @chargeairh · Token: $CHARGE 0x7b8892cea949bc60d494681d1bc3ea96071ab0b8

surplus-launcher: 0x3ac82d45ff9e4963cefbc3f74dbe293f34c927ca

What a read looks like

User (the facts card the Charge bot and chargeai.bid build from a contract address):

What's the charge on this?
$SEA · Seaweed Token · 0x813d…35ba
age 9d17h · mc $72k · ATH $456k (-84%, 9d17h ago) · launched at $451k · direct launch, Uniswap v3 pool, non-ETH pair
chart 3d (3h bars): ▅▇▇▇▇▆███▇▃▁▃▂▂▁▂▂▂▂▂▂▂▁
range 24h $64k–$136k (now at 12% of it) · 72h $57k–$202k
since the ATH: low $57k (13% of ATH, 40h06m ago, tapped 2x) · now 1.3x off it
higher lows 0 · lower highs 1 (6h blocks) · volatility 8%/h
moves: 1h -6% · 6h -12% · 24h -4% · 3d -60%
last 1h: 94 trades · 29 buyers / 31 sellers · buys $427 / sells $794 · -6%
last 6h: 969 trades · 112 buyers / 118 sellers · buys $6.6k / sells $7.5k · -12% · volume 5.0x the 6h before · buyers 2.4x the 6h before
last 24h: 3141 trades · 273 buyers / 285 sellers · buys $26k / sells $25k · -4% · volume 0.9x the day before · buyers 1.0x the day before
wallets: 80 new in 6h · 209 new in 24h · holders ~360 (+2 in 24h) · 8% of the buyers from 24-48h ago still hold
cost basis: 10% of held supply in profit · 64% between break-even and 2x underwater
since launch: 2405 buyers · 2373 sellers · median buy $14 · top 3 buyers = 12% of buys
top 10 hold 39% · biggest 14.4%
dev: holds 0.8% · bought $2.5k · sold $0
first-30s buyers: 577, bought over 100% of supply (churned) · sold 96% of it
creator: 0 other launches in the 24h before
ticker: 0 other $SEA launched in the 24h before
socials: unknown

Charge:

⚡ Charge 82% · charged
Chart: 84% under its ATH from 9d17h ago · sitting on its post-ATH low (tapped 2x) · 80 new wallets in 6h.
Next 24h, tokens I read like this: +50% before -33%: 35% · closed higher: 25% · new ATH: 3% · fell 50%: 35%.

What stands out:
- top 10 wallets hold 39%
± ~360 holders (moves both ways)
+ 2.4x the buyers of the 6h before
+ 80 new wallets in the last 6h

Levels: 24h high $136k · ATH $456k · post-ATH low $57k

Big downside in reads like this. Small size or no size. Not advice.

• Charge % ranks the setup among tokens that are trading at that moment (its odds of +50% before -33%); bands are low / mid / charged / overcharged. Higher charge means more action, not more safety.
• Odds are for the next 24 hours from the snapshot price: +50% before -33%, closed higher, a new ATH, fell 50% at some point; direct launches under a day old also get the rug rate (liquidity pulled).
• What stands out lines are fixed checks shown only when their direction was measured on the training days, within the same charge band and age group. Levels are prices the chart already reacted to, not targets.

How it was built

• Data: public trades on Robinhood Chain from 2026-10-05 to 2026-10-15: 25,153 tokens, snapshots at fixed ages and through their life while they traded (up to 10 days), each labelled with what happened over the next 24 hours. Public facts only (price, flows, buyers and sellers, holders estimated from trades, the creator's own trades, first-30s buyers, socials, and chart facts: ranges, the low since the ATH and its taps, higher lows, volume and new wallets vs the day before, retention, cost basis).
• Teacher: calibrated gradient-boosted models per outcome, every token weighted equally. Every answer is rendered from data by templates: no text written by another LLM.
• Fine-tune: Qwen/Qwen3-30B-A3B-Instruct-2507 with LoRA (rank 32, alpha 64) on attention and experts, 52,595 rows (2,149 general chat turns so it still talks like a model), 2 epochs, merged into these weights.

The exam

On a twin trained only on labels that ended before 2026-10-13 (charge-ai/charge-rh-30b-v3-test), scored on 2,891 held-out snapshots from 2026-10-13 to 2026-10-15 against what really happened (token-weighted AUC: 0.5 is a coin flip). Format: 100% of answers well-formed.

Main read (+50% before -33%), the same 600 snapshots for every reader, given the same facts:

Reader	All ages	Under a day old	A day or older
Charge v3	0.80	0.84	0.71
Claude Opus 5.5	0.76	0.75	0.75
Claude Sonnet 5.5	0.72	0.70	0.72
Charge 30B (v2)	0.74	0.79	0.65

Against Charge 30B on all 2,891 snapshots: +0.081 (95% interval +0.064 to +0.100) overall and +0.080 (+0.043 to +0.120) on tokens a day or older; a new ATH within 24 h 0.85 vs 0.61 and fell 50% 0.72 vs 0.66 on those older tokens. Where Claude leads: on tokens a day or older Opus 5.5 is ahead on the main read (0.75 vs 0.71) and both Claude models are better at "closed higher". The release model in this repo was trained on all days through 2026-10-15 (no held-out days left), the same recipe.

Limits

• It reads the facts it's given. It doesn't see contracts (honeypots and hook traps need a sell simulation), the team, the news or who is about to post. Holder counts in the card are estimated from trades and run high.
• It learned from 11 days of one chain. It is wrong often. It is a filter and a second read, not a buy button.
• Not financial advice.

Use

Send the facts card as the user message; the default system prompt is built into the chat template. Served on the Surplus API, and through chargeai.bid, the Charge Telegram bot and an OpenAI-compatible API (api.chargeai.bid/v1).

License

Apache-2.0, as the base model (Qwen/Qwen3-30B-A3B-Instruct-2507). See LICENSE.