Kronos vs TimesFM 2.5

Kronos is the first open-source foundation model purpose-built for financial candlestick (K-line / OHLCV) data, accepted at AAAI 2026. Instead of treating price series like text or images, Kronos uses a custom two-stage architecture: a specialized tokenizer that converts continuous OHLCV data into discrete tokens, followed by an autoregressive Transformer trained on data from 45+ global exchanges. Four model sizes range from 4.1M to 499M parameters, all released under MIT license. The model learns the statistical structure of market microstructure directly from raw candlestick sequences, enabling zero-shot and few-shot forecasting across asset classes — equities, crypto, and commodities. It ships with a live BTC/USDT prediction demo, Qlib integration for A-Share markets, and a backtesting framework so researchers can evaluate strategies end-to-end. With 13.6k GitHub stars in a niche domain, the community reception has been unusually strong. Kronos matters because most "AI for trading" projects glue LLMs to news sentiment or financial reports — pattern-matching on text rather than market structure. Kronos is the rare project that treats price action itself as the primary modality, giving quants and ML researchers a base model they can fine-tune on proprietary data rather than starting from scratch on every new dataset.

TimesFM 2.5 is the latest update to Google Research's pretrained time-series foundation model — a 200M parameter decoder-only model that does zero-shot forecasting across virtually any time-series domain without needing to retrain or fine-tune. Released March 31, 2026, it expands context length to 16,000 time steps (up from earlier versions) and adds an optional 30M continuous quantile head for probabilistic forecasting up to 1,000 steps ahead. Unlike traditional forecasting approaches that require training a new model per dataset, TimesFM was pre-trained on 100 billion real-world time points across diverse domains. You point it at new data — retail sales, server metrics, energy demand, financial prices — and it forecasts without any additional training. The March 31 update also restores covariate (XReg) support and updates inference APIs for better integration. With 14,000 GitHub stars and trending today, TimesFM is becoming the default baseline for time-series work in the same way BERT became the baseline for NLP tasks. Google Cloud users get it directly via BigQuery ML's AI.FORECAST function. For everyone else, it's available on HuggingFace and installable as a Python package.