YingLong 300M

YingLong 300M is the largest checkpoint in the YingLong family, offering the highest forecast quality at the cost of increased latency. The official paper frames it as the top-capacity YingLong release for zero-shot forecasting, the ceiling of the family rather than its cheapest tier.

The broader family continues to use direct quantile-style probabilistic outputs rather than a language-model-style decoder, and YingLong 300M keeps that non-causal transformer design: bidirectional attention over the input history feeding a multi-quantile output head, so calibrated prediction intervals come straight out of a single forward pass. The official model card states the released YingLong checkpoints were pre-trained on 78B time points, the same corpus that gives the family its zero-shot reach across domains.

On TSFM.ai reach for YingLong 300M when accuracy is the primary concern and the workload can absorb the higher latency that comes with the largest checkpoint. Step down to YingLong 110M when you want most of the quality at lower cost, to 50M for a balanced default, or to 6M when throughput and per-call latency dominate.