Kronos
A free AI model trained specifically on stock and crypto price charts (rather than general text) that tries to predict where a market might move next, similar to how a language model predicts the next word.
🔗 Visit Kronos
Description
Most AI models are trained on text, code or images — none of that is a great fit for predicting financial markets, where the real signal is in price movement patterns (candlestick charts) over time. Kronos takes a different approach: it's trained from the ground up on actual price-chart data from dozens of exchanges, treating market movements the way a language model treats words, learning to predict what typically comes next in a price sequence.
Kronos is a free, open-source (MIT license, ~33,500 GitHub stars) foundation model — the first of its kind published openly — trained on K-line (candlestick) data from 45+ global exchanges, using a decoder-only Transformer architecture with hierarchical discrete tokenization of price/volume data. It comes in four sizes (mini at 4.1M parameters up to large at 499M parameters), includes fine-tuning scripts for custom datasets, and has a live public demo forecasting BTC/USDT price movements 24 hours out. The underlying research paper was accepted by AAAI 2026, a top-tier AI conference.
💬 Our review
The short version: if you're a quant researcher or serious trader who wants an open, fine-tunable foundation model built specifically for price data instead of adapting a general-purpose LLM, Kronos is a legitimate, peer-reviewed research project worth experimenting with — but it is not a magic market-prediction tool.
Being accepted at AAAI 2026 gives it real academic credibility that most "AI trading" GitHub projects lack, and offering four model sizes with fine-tuning scripts makes it genuinely usable rather than just a research artifact. The critical caveat, true of literally every financial forecasting model ever built: past price patterns predicting future price patterns is a notoriously hard problem, and no public benchmark here should be mistaken for real trading edge — markets are adversarial and any publicly known pattern tends to get arbitraged away quickly. Treat this as a serious research starting point for building your own models, not as a signal generator to trade on directly.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Entièrement gratuit et open source (licence MIT), modèles disponibles sur Hugging Face.
Pros
Gratuit et open source, article accepté à AAAI 2026 (conférence IA de premier plan)
4 tailles de modèle, scripts de fine-tuning fournis
Entraîné spécifiquement sur des données de 45+ bourses mondiales
Démo publique en direct (prévision BTC/USDT)
Cons
Aucun modèle de prédiction de marché, aussi sophistiqué soit-il, ne garantit un avantage de trading réel
Marché financier adversarial : tout signal public connu tend à s'arbitrer rapidement
Nécessite des compétences en ML pour être utilisé sérieusement (pas un outil clé en main pour trader)
Projet de recherche récent (papier août 2025), peu de recul en usage réel