YingLong 6M

YingLong 6M is the smallest checkpoint in the YingLong family, built for efficient zero-shot probabilistic forecasting. The official card and released code position it as a lightweight YingLong checkpoint, and at 6M parameters it is the most latency-friendly YingLong option in the catalog — the entry point to the family for anyone who wants foundation-model forecasting at the lowest possible serving footprint.

Architecturally it is a non-causal transformer forecaster: rather than the autoregressive decoder used by language-model-style forecasters, it relies on bidirectional attention over the input history and a multi-quantile output head, so a single forward pass emits calibrated quantiles instead of a sampled point trajectory. Like the rest of the released YingLong checkpoints, the official model card states it was pre-trained on 78B time points, which is what gives the family its broad zero-shot reach.

On TSFM.ai reach for YingLong 6M when throughput and per-call latency dominate — running many series in parallel, or any workload where the cheapest probabilistic forecaster that still generalizes zero-shot is the right tradeoff. Step up to YingLong 50M when you want more capacity without much added latency, and to YingLong 110M or 300M when forecast quality justifies the heavier checkpoints.