YingLong 50M

YingLong 50M is the mid-size checkpoint in the YingLong family, balancing forecast quality against inference speed. It is the natural default in the family: more capacity than the 6M variant for cases where the smallest checkpoint leaves accuracy on the table, but lighter than the 110M and 300M models when their latency would be hard to justify.

Like the smaller YingLong release, the official first-party surfaces describe a bidirectional forecasting model with quantile outputs rather than a causal decoder — a non-causal transformer that attends over the full input history and emits calibrated quantiles directly from a multi-quantile output head. The official model card states the released YingLong checkpoints were pre-trained on 78B time points, the same pretraining corpus that underpins the family's zero-shot generalization.

On TSFM.ai reach for YingLong 50M when you want a sensible balanced default for zero-shot probabilistic forecasting. Drop to YingLong 6M when throughput or per-call latency is the binding constraint, and step up to YingLong 110M or 300M when a workload rewards the extra forecast quality enough to absorb the higher latency.