Monte

The model itself is kind of wild. It fits in 65 bytes. That is it. The entire neural network lives in 65 bytes.

For comparison, this tweet alone is around 280 bytes, a typical PyTorch model ranges from 50 to 500 MB, and GPT-4 is estimated at roughly 1.7 TB.

Trained on 9,012 samples of SOL/USDT (Binance 1-min candles).
Tested on 1,803 unseen samples.
59.2% accuracy. 30-minute lookahead. Better than a coin flip.

But that is not the point. The point is that a neural network can live and compute entirely on-chain, fully verifiable, trustless, and immutable.

Repo for the experiment if you want to dig in:
github.com/QuantuLabs/pri…

Since this kind of experiment has never been done on solana before, I'd recommend validators keep an eye on their servers next week. Odds are everything will be fine, but still.

Permissionless perps for long-tail assets are coming to Solana.

Next week we unleash the experiment.

One thing that always bothered me about perps was how limited market creation still feels.

Having to bootstrap a discovery pool for every new market, using the same memecoin as collateral to trade it with leverage, doesn't sound flexible.

Playing around with Percolator got me thinking about what else perps could be. There's already so much innovation happening on Solana, yet perps still feel way too close to TradFi.

That's why I went with Mimir, our fully onchain oracle, instead of relying on a dedicated discovery pool for every market.

The idea goes way beyond trading the usual assets. Any project should be able to launch its own perp market, and traders should have access to everything from long-tail tokens to entirely new indices.

We've got everything we need to make TradFi perps look outdated. Time to prove it.