Kronos open-sources market foundation model

This is one of the more credible finance-model releases in a while because it leans into a domain-specific representation instead of pretending generic time-series models are enough. The real question is not whether the paper is ambitious, but whether the reported gains survive live trading assumptions, costs, and regime shifts.

• –The tokenizer idea matters: discretizing OHLCV into tokens gives the model a market-native representation rather than raw numeric sequences.
• –The scale is the headline here: 12B K-lines across 45 exchanges is enough to justify the foundation-model framing.
• –The paper’s strongest claim is zero-shot transfer, but that is also where finance models often break once you add slippage, fees, and nonstationarity.
• –Synthetic K-line generation and volatility forecasting make Kronos more interesting than a pure price-prediction model.
• –Early community reaction is already cautious, which is healthy for a finance model that claims broad generalization without task-specific training.