WEIGHTS.FUND @weights_fund SOLANA
Launchpad for open-source AIPre-launch · nothing deployed

Fees
become
weights.

Launch a token for an open model. Every trade pays a 1% fee — 60% of it buys GPU time for that model, released milestone by milestone. When the run ends, the weights, logs and evals are published for everyone.

Launch a model token → Browse runs
Launch date and CA will be posted only on @weights_fund →
1%fee per trade
0.60%to GPU vault
3paid milestones
100%open weights
RUN · tiny-uk-1b · 8× H100SIMULATED
Step0
Loss—
Tokens/s—
GPU-h used0
VAULT → GPU INVOICE #0007EXAMPLE DATA
1%trading fee on every launched token
60%of that fee can only be spent on GPU invoices
M1→M3compute released per milestone, never upfront
0models trained so far — we're pre-launch
01 / HOW IT WORKS

Four steps.
One output.

A token on weights.fund is attached to a model project — a written spec of what will be trained, on which data, and under which license. Trading funds the spec. The spec produces weights.

01

Launch

Create a token and attach a model spec: architecture, size, dataset, license, deadline.

  • ~0.02 SOL launch fee
  • spec frozen at launch
  • open license required
02

Trade

Every buy and sell pays 1%. The compute share goes into the project's vault, not to anyone's wallet.

  • 0.60% → compute vault
  • 0.25% → maintainer
  • 0.15% → platform
03

Train

The vault pays GPU providers directly, in tranches, against milestones. Logs stream in public.

  • invoices paid from vault
  • live loss curves
  • checkpoints hashed
04

Publish

Weights, config, training logs and eval scripts are released. The weights' SHA-256 is recorded on-chain.

  • Apache-2.0 / MIT / CC-BY
  • reproducible evals
  • fees fund v2 next
02 / FEE SPLIT

Where the
1% goes.

Planned parameters for v1. The compute share is the point of the whole thing, so it's the biggest slice and the most restricted one.

Compute vault0.60%
Maintainer0.25%
Platform0.15%

Compute vault · 60%

Held per project. Can only pay invoices from approved GPU providers for that model's spec. Released per milestone. Unspent compute after a missed deadline rolls into the Community Queue.

Maintainer · 25%

For the person or team doing the work: data cleaning, training, evals, writing it up. Streams continuously, so maintainers aren't forced to pump for income.

Platform · 15%

Runs the launchpad, the indexer, the proof registry and audits. No other platform fee, no hidden spread.

Simulate: total trading volume
$250,000
Slide to see what a given volume would fund at $2.00 per H100-hour (example rate).
Compute vault$1,500
H100-hours750
Enough for~1B model
03 / LAUNCH

Launch a
model token.

Fill in the token, then the model spec. The calculator estimates the compute your spec needs and how much trading it would take to fund it. Launches open after audit — you can prepare yours now.

Name needs at least 2 characters.
2–10 letters or digits.
Image
0/280 · shown on the token page
Describe the model in at least 20 characters.
Project type
Est. GPU-hours84
Est. compute$169
+ buffer ×3$506
Volume to fund$84k
6·N·D FLOPs on H100 at 40% utilisation, $2.00/GPU-h. The ×3 buffer covers experiments, failed starts and evals — it becomes the vault goal.
Milestone payout
M1 data card / M2 training / M3 weights
Dev tokens are locked for 6 months automatically.
Visibility
Launch fee0.02 SOL
Dev buy0 SOL
Total0.02 SOL
04 / RUNS

The board.

How projects will look once launches open. These are example projects with made-up numbers — click one to see its milestones and proof receipt.

No runs match. Try another filter.

EXAMPLE PROJECTS · NOT REAL TOKENS · NO REAL RUNS

05 / COMPUTE CALCULATOR

What does
a model cost?

The rule of thumb every lab uses: training takes about 6 × parameters × tokens FLOPs. Pick a size, a GPU and a price. Real runs cost more — this gives you the order of magnitude.

Parameters1B
Training tokens20B
20 tokens / param
GPU
$ / GPU-hour
GPUs
Utilisation (MFU)40%
GPU-hours84
Wall-clock10.5 h
Compute cost$169
Trading volume to fund it$28k
6 × 1e9 × 2e10 = 1.20e20 FLOPs
Peak dense BF16 throughput per GPU. Prices are examples — change them. Not a quote.
Share this estimate on X
06 / PROOF

Don't trust
the curve. Download
the weights.

Every milestone produces something you can check yourself. The registry stores hashes on-chain; the files live on public model hubs and repos.

ArtifactWhat's publishedHow you verify itMilestone
Data cardSources, licenses, filtering code, token counts, dedup stats.Re-run the filter script, compare counts.M1
GPU invoicesEvery invoice the vault paid: provider, hours, GPU type, amount.Vault transactions match invoice totals.M2
Training logsLoss, learning rate, throughput, every eval step, streamed live.Logs line up with GPU-hours billed.M2
WeightsFinal checkpoint in safetensors + config + tokenizer.SHA-256 of the file equals the on-chain record.M3
EvalsBenchmark scores with the exact scripts and seeds.Run the script on the weights, get the same numbers.M3
Try it · hash a file locally

Check a weights file

Pick any file. Your browser computes its SHA-256 right here — nothing is uploaded. On launch, you'd compare it with the hash recorded on-chain.

SHA-256 will appear here.
Compare with a registry record

Match the record

Records shown are examples from the demo board.
07 / MILESTONE ESCROW

No weights,
no next payout.

The vault never hands a lump sum to anyone. It pays GPU providers in tranches, and each tranche needs the previous milestone's proof.

M0
0%

Launch

Spec frozen: model size, data, license, deadline. Fees start filling the vault.

M1
10%

Data card

Dataset and filtering published. Unlocks compute for data processing.

M2
75%

Training

Paid in tranches against invoices. Logs and checkpoints published as it runs.

M3
15%

Weights

Weights + evals released, hash on-chain. Later fees fund a v2 run.

✕
Missed deadline

Roll over

Unspent compute moves to the Community Queue — other open projects, chosen by holders' vote.

PLANNED DESIGN · SUBJECT TO AUDIT · NOT DEPLOYED

08 / STATUS

Where things
actually are.

This list gets updated before anything else on this site does.

Site + launch flowDone
Compute calculatorDone
Local hash verifierDone
Vault + milestone programIn progress
GPU provider integrationNot started
Proof registryNot started
AuditNot started
Models trained0
09 / FAQ

Questions.

10 / FOLLOW THE RUNS

Updates
live on X.

Build status, new projects, milestone proofs and the contract address on launch day — posted only from @weights_fund. We never DM first and never run giveaways.

Follow @weights_fund → Share weights.fund
01CA only from @weights_fund, only as text.
02No DMs. No presale. No airdrop.
03Status list posted every time it changes.
weights.fund avatar
weights.fund
@weights_fund
Follow

Fees become weights. A launchpad where trading fees buy GPU time for open-source AI models. No weights, no next payout. Pre-launch, nothing deployed.

📍 the training logs

📌 Pinned

Fees become weights.

weights.fund is a launchpad for open-source AI. Every token funds one model project.

1% fee per trade → 0.60% to that model's compute vault → paid to GPU providers, per milestone → weights, logs and evals published openly.

Open on Xx.com/weights_fund →
What the account posts
Thread

Fees become weights. A launchpad where every token is attached to an open-source model — and trading fees pay for the GPUs that train it.

 The math

1B params × 20B tokens × 6 = 1.2e20 FLOPs → ≈ 84 H100-hours → ≈ $170 at $2/hour. Real runs cost ~3× that. Launches budget for it.

 Status

Site + launch flow — done. Vault program — in progress. GPU integration, audit — not started. Models trained — 0.