Kronos — financial-market foundation model
Billed as "the first open-source foundation model for the language of financial markets." A domain-specific financial time-series model — not a general LLM — it is the prediction layer of this wiki's founding trio: where TensorTrade learns a policy and banbot executes orders, Kronos forecasts the market the other two act on. T3: an author-published open-weight release with a clear method, but its forecasting accuracy is self-reported.
What it is
- Modality. Operates on K-line / candlestick (OHLCV) data — open/high/low/close/volume/amount — the native language of price history, not natural-language text.
- Tokenizer. A specialized tokenizer quantizes the continuous, multi-dimensional candlestick series into hierarchical discrete tokens.
- Model. A decoder-only autoregressive Transformer then forecasts future OHLCV bars — the same tokenize-then-autoregress recipe behind text LLMs, retargeted to financial series and engineered for the high-noise character of market data.
- Training & release. Pre-trained on 45+ global exchanges. MIT open-weight, three sizes on Hugging Face: Kronos-mini (4.1M), small (24.7M), base (102.3M).
Lineage & the name clash
Kronos sits in the time-series foundation model
sub-field alongside Amazon Chronos, Google TimesFM, and Salesforce Moirai — note the
Kronos vs Chronos near-collision. It shares the open-weight foundation-model pattern with the
text-LLM market (../llm-providers-wiki open-weight-models: tokenizer → autoregressive
Transformer), but the OHLCV modality and absence of a provider-market angle keep it here, not there.
Why it matters here
Kronos was the first financial-ml source parked in the hub _inbox — the seed of the cluster
this spoke grew from. It anchors the predict corner of the predict → learn → execute stack.
Re-verified 2026-08-30 (quality cycle) — big, and stopped
Stale at 62 days; re-read against the repository. 38,101★ / 6,342 forks, 270 open issues, MIT, not archived (GitHub API, 2026-08-30) — by a wide margin the most-starred source in this spoke, and one of the largest in the whole corpus. Last push 2026-04-13, four and a half months ago.
Those two facts belong together, because this spoke's founding source turns out to be a repository with enormous adoption and no recent maintenance, carrying 270 open issues nobody is closing. It does not retract anything above — the model weights and the method are what they were — but it changes what "open-weight release with a clear method" is worth to a reader deciding whether to build on it. The self-reported forecasting accuracy noted above has still never been independently re-run, and now there is nobody upstream to ask.
Related
financial-time-series-foundation-model · algorithmic-trading · tensortrade · banbot · synthesis