Time-MoE

Time-MoE, published in September 2024 and presented at ICLR 2025, was the first time-series foundation model to scale with a sparse mixture-of-experts design. The largest variant totals 2.4 billion parameters but activates only around 24 million per time step, pretrained on Time-300B — 300 billion time points of diverse series. Unlike tokenized forecasters in the Chronos style, it regresses directly on scaled values with a point-forecast head, avoiding quantization error.

The paper reported that Time-MoE matched task-specific state-of-the-art on long-term forecasting with zero shots, and third-party surveys noted it still held best numbers on several ETTh1/ETTm1 settings at horizon 96 as of late 2025. It's an academic artifact rather than a maintained product — the reference repo is a research codebase — but as an Apache 2.0 proof that sparse scaling beats dense scaling for forecasting, it remains influential. Sizes run 50M to 2.4B on Hugging Face.

• Reported in the Time-MoE paper (arXiv 2409.16040): competitive with task-specific SOTA on long-term forecasting benchmarks with zero-shot inference (author-reported)
• Third-party survey notes (2025-26): still holds best numbers on several ETTh1/ETTm1 settings at horizon 96
• No canonical independent benchmark table verified