Mnemonic Labs is training a foundation model for quant trading: Reveneau

Why it mattersA trading desk that paid two quants to write strategies is a market this team is pointing at. Buyers of quant work can test a model that writes its own.

A trading desk writes strategies by hand, one at a time. Two brothers have launched a company that trains one model to write all of them.

Mnemonic Labs launched on Y Combinator's launch page, carrying 27 votes so far. The company's YC batch is Fall 2026. The pitch: a foundation model for quant trading that learns strategies directly from market activity, rather than from hard-coded rules.

What the team is building

The founders say the model is built on one bet. Trading is a dynamic domain, like protein folding or self-driving cars. Text is not the most useful way to represent the data. The team says a model trained end to end on markets will find strategies that hard-coded priors would miss.

The company lists numbers for its own progress. It built its training pipeline and model architecture in less than six weeks. Its current training run uses over 25 TB of tokens. The company names no customers, no revenue and no investors on its launch page.

Who the founders are

Mnemonic Labs says Videh and Kalki Seksaria are brothers and both MIT graduates. The company says Videh was the youngest quant at Citadel at 19. It says Kalki made 20 million US dollars in profit at Tower Research at 22. The company says the two are five-time USACO finalists combined.

All of those are the company's own claims, from its launch page. Citadel and Tower Research have not confirmed any of them, and the company names no source outside itself.

Why they think the market is open

The company's own account is that large trading firms optimise for short-term profit and do not take research-heavy bets. New trading desks have to be profitable in a year or they are closed. Already profitable desks do not want to gamble on new approaches.

The founders also think foundation labs will not enter this market. Their argument: a lab like Google could run the world's best hedge fund if it wanted to. But it will choose to sell to many other markets instead. Trading is one of the few domains where foundation labs will not compete directly.

What a buyer should know

Nothing is public about the model's performance. A buyer has nothing independent to compare against yet. The launch page does not say how the model will be sold. It may be offered as a service, licensed to a trading firm, or kept in-house. The 25 TB figure is a training data size. It is not a measure of how well the model trades. A trading firm thinking about a deal should ask for a track record first. The record must be on a real account, under the buyer's risk limits, over a measured window.

Source

Mnemonic Labs: Training a Foundation Model for Quant Trading, Y Combinator launches.
Mnemonic Labs, company website.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.