[ { "name": "EarthPT", "family": null, "org": "Aspia Space / U. Hertfordshire", "release_date": "2023-09-13", "paper_arxiv_id": "2309.07207", "paper_url": "https://arxiv.org/abs/2309.07207", "params": "700M", "weights": "open", "architecture": "decoder-only transformer (GPT-like)", "task_focus": "Earth observation (multi-spectral satellite time series)", "notes": "First Earth Observation TSFM; autoregressive pixel-level reflectance forecasting." }, { "name": "TimeGPT-1", "family": "TimeGPT", "org": "Nixtla", "release_date": "2023-10-05", "paper_arxiv_id": "2310.03589", "paper_url": "https://arxiv.org/abs/2310.03589", "params": "unknown", "weights": "closed-api", "architecture": "encoder-decoder transformer with self-attention", "task_focus": "general forecasting + anomaly detection", "notes": "First production-ready commercial TSFM API; trained on 100B+ data points across many domains." }, { "name": "Lag-Llama", "family": null, "org": "ServiceNow Research / Mila / U. Montreal", "release_date": "2023-10-12", "paper_arxiv_id": "2310.08278", "paper_url": "https://arxiv.org/abs/2310.08278", "params": "2.45M", "weights": "open", "architecture": "decoder-only transformer (LLaMA-style with RoPE, RMSNorm) using lag covariates", "task_focus": "general univariate probabilistic forecasting", "notes": "First fully open foundation model framed as 'LLaMA for time series'; small (~2.45M params) but strong zero-shot." }, { "name": "TimesFM (PreDcT, v1.0)", "family": "TimesFM", "org": "Google Research", "release_date": "2023-10-14", "paper_arxiv_id": "2310.10688", "paper_url": "https://arxiv.org/abs/2310.10688", "params": "200M", "weights": "open", "architecture": "decoder-only transformer with patched input", "task_focus": "general forecasting", "notes": "Original paper; checkpoint timesfm-1.0-200m released on HuggingFace in Feb 2024 under Apache 2.0." }, { "name": "ForecastPFN", "family": null, "org": "Abacus.AI / U. Maryland", "release_date": "2023-11-03", "paper_arxiv_id": "2311.01933", "paper_url": "https://arxiv.org/abs/2311.01933", "params": "unknown", "weights": "open-research-only", "architecture": "prior-data fitted network (transformer encoder, Bayesian inference approximation)", "task_focus": "zero-shot forecasting on small samples", "notes": "First zero-shot forecaster trained purely on synthetic data; NeurIPS 2023." }, { "name": "Tiny Time Mixers (TTM-r1)", "family": "Granite TTM", "org": "IBM Research", "release_date": "2024-01-08", "paper_arxiv_id": "2401.03955", "paper_url": "https://arxiv.org/abs/2401.03955", "params": "1M-5M", "weights": "open", "architecture": "TSMixer-based MLP (adaptive patching, resolution prefix tuning)", "task_focus": "general multivariate forecasting (CPU-friendly)", "notes": "Compact pre-trained mixer; NeurIPS 2024; first 'tiny' TSFM that can run on CPU." }, { "name": "Moirai 1.0", "family": "Moirai", "org": "Salesforce AI Research", "release_date": "2024-02-04", "paper_arxiv_id": "2402.02592", "paper_url": "https://arxiv.org/abs/2402.02592", "params": "14M / 91M / 311M", "weights": "open-research-only", "architecture": "masked encoder-only transformer with multi-patch projections", "task_focus": "universal probabilistic forecasting", "notes": "Trained on LOTSA archive (27B observations across 9 domains); released in Small/Base/Large." }, { "name": "Timer", "family": "Timer", "org": "Tsinghua University (THUML)", "release_date": "2024-02-04", "paper_arxiv_id": "2402.02368", "paper_url": "https://arxiv.org/abs/2402.02368", "params": "~84M", "weights": "open", "architecture": "decoder-only transformer (autoregressive)", "task_focus": "general forecasting, imputation, anomaly detection", "notes": "ICML 2024; pretrained on 260B time points; 'GPT for time series'." }, { "name": "MOMENT", "family": "MOMENT", "org": "CMU / U. Penn (Auton Lab)", "release_date": "2024-02-06", "paper_arxiv_id": "2402.03885", "paper_url": "https://arxiv.org/abs/2402.03885", "params": "40M / 125M / 385M", "weights": "open", "architecture": "encoder-only T5-based transformer with patched masked reconstruction", "task_focus": "general analysis (forecasting, classification, anomaly detection, imputation)", "notes": "ICML 2024; introduced the Time Series Pile pretraining corpus; Small/Base/Large variants." }, { "name": "TimesFM-1.0-200m", "family": "TimesFM", "org": "Google Research", "release_date": "2024-02-12", "paper_arxiv_id": "2310.10688", "paper_url": "https://huggingface.co/google/timesfm-1.0-200m", "params": "200M", "weights": "open", "architecture": "decoder-only transformer with patched input", "task_focus": "general forecasting", "notes": "First open checkpoint released to HuggingFace; context up to 512, any horizon." }, { "name": "GTT (General Time Transformer)", "family": null, "org": "Lenovo / Independent", "release_date": "2024-02-12", "paper_arxiv_id": "2402.07570", "paper_url": "https://arxiv.org/abs/2402.07570", "params": "unknown", "weights": "open", "architecture": "encoder-only transformer (next curve-shape prediction)", "task_focus": "zero-shot multivariate forecasting", "notes": "CIKM 2024; pretrained on 200M curve-shape samples; channel-wise reformulation." }, { "name": "UniTS", "family": null, "org": "Harvard (Zitnik Lab) / MIT-IBM Lincoln Lab", "release_date": "2024-02-29", "paper_arxiv_id": "2403.00131", "paper_url": "https://arxiv.org/abs/2403.00131", "params": "unknown", "weights": "open", "architecture": "modified transformer with task tokenization", "task_focus": "unified multi-task (forecasting, classification, imputation, anomaly)", "notes": "NeurIPS 2024; single architecture for predictive + generative TS tasks." }, { "name": "Chronos", "family": "Chronos", "org": "Amazon (AWS AI Labs)", "release_date": "2024-03-12", "paper_arxiv_id": "2403.07815", "paper_url": "https://arxiv.org/abs/2403.07815", "params": "8M / 20M / 46M / 200M / 710M", "weights": "open", "architecture": "T5 encoder-decoder, language-model-style with quantized tokens", "task_focus": "general probabilistic forecasting", "notes": "Tokenize-and-LM approach; tiny/mini/small/base/large variants." }, { "name": "Moirai 1.1", "family": "Moirai", "org": "Salesforce AI Research", "release_date": "2024-06-18", "paper_arxiv_id": "2402.02592", "paper_url": "https://huggingface.co/Salesforce/moirai-1.1-R-large", "params": "14M / 91M / 311M", "weights": "open-research-only", "architecture": "masked encoder-only transformer", "task_focus": "universal probabilistic forecasting", "notes": "Iterative update of Moirai 1.0 with ~20% improvement on low-frequency (yearly/quarterly) data." }, { "name": "Toto", "family": "Toto", "org": "Datadog", "release_date": "2024-07-10", "paper_arxiv_id": "2407.07874", "paper_url": "https://arxiv.org/abs/2407.07874", "params": "151M", "weights": "open", "architecture": "decoder-only transformer with factorized space-time attention; Student-T mixture output", "task_focus": "observability metrics + general forecasting", "notes": "First TSFM tuned for observability metrics; trained on 1T points (75% Datadog telemetry)." }, { "name": "VisionTS", "family": null, "org": "Yonsei / Tsinghua", "release_date": "2024-08-30", "paper_arxiv_id": "2408.17253", "paper_url": "https://arxiv.org/abs/2408.17253", "params": "unknown (uses MAE-ViT backbone)", "weights": "open", "architecture": "vision MAE (masked autoencoder) repurposed as time series forecaster", "task_focus": "zero-shot univariate forecasting", "notes": "Free-lunch transfer from vision: renders TS as grayscale image for MAE reconstruction." }, { "name": "TimeDiT", "family": null, "org": "USC / Microsoft", "release_date": "2024-09-03", "paper_arxiv_id": "2409.02322", "paper_url": "https://arxiv.org/abs/2409.02322", "params": "unknown", "weights": "open-research-only", "architecture": "diffusion transformer (DiT) with unified masking", "task_focus": "general forecasting + imputation + missing-value handling", "notes": "ICML 2024 workshop; first general-purpose diffusion-transformer TSFM with editing capability." }, { "name": "Time-MoE", "family": "Time-MoE", "org": "Princeton / ECNU / collaborators", "release_date": "2024-09-24", "paper_arxiv_id": "2409.16040", "paper_url": "https://arxiv.org/abs/2409.16040", "params": "up to 2.4B (sparse MoE)", "weights": "open", "architecture": "decoder-only transformer with sparse mixture of experts", "task_focus": "general forecasting (autoregressive)", "notes": "ICLR 2025 Spotlight; first billion-scale TSFM; introduced Time-300B dataset." }, { "name": "Timer-XL", "family": "Timer", "org": "Tsinghua University (THUML)", "release_date": "2024-10-07", "paper_arxiv_id": "2410.04803", "paper_url": "https://arxiv.org/abs/2410.04803", "params": "unknown", "weights": "open", "architecture": "decoder-only transformer with TimeAttention (long-context generalization)", "task_focus": "univariate/multivariate/covariate-informed forecasting", "notes": "ICLR 2025; extends Timer with thousands-token context and multivariate next-token prediction." }, { "name": "Moirai-MoE", "family": "Moirai", "org": "Salesforce AI Research", "release_date": "2024-10-14", "paper_arxiv_id": "2410.10469", "paper_url": "https://arxiv.org/abs/2410.10469", "params": "small/base variants (sparse MoE)", "weights": "open-research-only", "architecture": "decoder-only transformer with sparse mixture of experts", "task_focus": "universal forecasting", "notes": "ICML 2025; first MoE TSFM; +17% over Moirai 1.0 with 65x fewer activated params than Chronos/TimesFM." }, { "name": "Chronos-Bolt", "family": "Chronos", "org": "Amazon (AWS AI Labs)", "release_date": "2024-11-26", "paper_arxiv_id": null, "paper_url": "https://huggingface.co/amazon/chronos-bolt-base", "params": "9M / 21M / 48M / 205M", "weights": "open", "architecture": "T5 encoder-decoder with patched direct multi-step forecasting", "task_focus": "fast general forecasting", "notes": "Patch-based variant of Chronos; ~5% more accurate, 250x faster, 20x more memory-efficient." }, { "name": "Granite TTM r2 / r2.1", "family": "Granite TTM", "org": "IBM Research", "release_date": "2024-12-01", "paper_arxiv_id": "2401.03955", "paper_url": "https://huggingface.co/ibm-granite/granite-timeseries-ttm-r2", "params": "1M-5M", "weights": "open", "architecture": "TSMixer MLP", "task_focus": "multivariate forecasting (minutely to weekly)", "notes": "Trained on ~700M (r2) / ~1B (r2.1) samples; >15% gain over r1; r2.1 adds daily/weekly resolutions." }, { "name": "TimesFM 2.0 (500m)", "family": "TimesFM", "org": "Google Research", "release_date": "2025-01-10", "paper_arxiv_id": "2310.10688", "paper_url": "https://huggingface.co/google/timesfm-2.0-500m-pytorch", "params": "500M", "weights": "open", "architecture": "decoder-only transformer with patched input", "task_focus": "general forecasting", "notes": "4x context (up to 2048) and ~25% better than v1; #1 on GIFT-Eval at release." }, { "name": "Sundial", "family": null, "org": "Tsinghua University (THUML)", "release_date": "2025-02-02", "paper_arxiv_id": "2502.00816", "paper_url": "https://arxiv.org/abs/2502.00816", "params": "unknown (family of sizes)", "weights": "open", "architecture": "transformer with TimeFlow loss (flow-matching, no tokenization)", "task_focus": "general probabilistic forecasting", "notes": "ICML 2025 Oral; flow-matching native continuous training on TimeBench (1T points)." }, { "name": "Mantis", "family": null, "org": "Huawei / collaborators", "release_date": "2025-02-21", "paper_arxiv_id": "2502.15637", "paper_url": "https://arxiv.org/abs/2502.15637", "params": "unknown (ViT-based)", "weights": "open", "architecture": "Vision Transformer (ViT) trained via contrastive learning", "task_focus": "time series classification", "notes": "Lightweight calibrated foundation model specifically for TS classification, not forecasting." }, { "name": "TimePFN", "family": null, "org": "Bilkent University", "release_date": "2025-02-22", "paper_arxiv_id": "2502.16294", "paper_url": "https://arxiv.org/abs/2502.16294", "params": "unknown", "weights": "open", "architecture": "prior-data fitted network for multivariate time series", "task_focus": "multivariate forecasting via synthetic data", "notes": "AAAI 2025; multivariate extension of ForecastPFN style." }, { "name": "TimeFound", "family": null, "org": "U. Illinois Urbana-Champaign", "release_date": "2025-03-06", "paper_arxiv_id": "2503.04118", "paper_url": "https://arxiv.org/abs/2503.04118", "params": "200M / 710M", "weights": "open-research-only", "architecture": "encoder-decoder transformer with multi-resolution patching", "task_focus": "general zero-shot forecasting", "notes": "Two sizes (200M/710M) trained on real + synthetic time series." }, { "name": "Toto-Open-Base-1.0", "family": "Toto", "org": "Datadog", "release_date": "2025-05-19", "paper_arxiv_id": "2505.14766", "paper_url": "https://arxiv.org/abs/2505.14766", "params": "151M", "weights": "open", "architecture": "decoder-only transformer with proportional factorized space-time attention", "task_focus": "observability metrics + general forecasting", "notes": "Open-weights release under Apache 2.0; trained on 2T+ data points (largest open TSFM corpus at the time); released with BOOM benchmark." }, { "name": "Kronos", "family": null, "org": "Tsinghua University", "release_date": "2025-08-04", "paper_arxiv_id": "2508.02739", "paper_url": "https://arxiv.org/abs/2508.02739", "params": "unknown (family)", "weights": "open", "architecture": "decoder-only transformer with hierarchical discrete tokenizer for OHLCV", "task_focus": "financial markets (K-line / OHLCV forecasting)", "notes": "First open-source TSFM purpose-built for financial candlestick data; trained on 12B+ K-line records from 45 exchanges." }, { "name": "FlowState", "family": null, "org": "IBM Research", "release_date": "2025-08-07", "paper_arxiv_id": "2508.05287", "paper_url": "https://arxiv.org/abs/2508.05287", "params": "9.1M", "weights": "open", "architecture": "state-space-model (SSM) encoder + functional basis decoder", "task_focus": "sampling-rate-invariant general forecasting", "notes": "NeurIPS 2025; smallest model in GIFT-Eval top 10; only TSFM that natively supports continuous-time inference." }, { "name": "VisionTS++", "family": null, "org": "Yonsei / Tsinghua", "release_date": "2025-08-06", "paper_arxiv_id": "2508.04379", "paper_url": "https://arxiv.org/abs/2508.04379", "params": "unknown (vision backbone)", "weights": "open", "architecture": "continual pretrained vision backbone with RGB multivariate encoding + multi-quantile heads", "task_focus": "cross-modal forecasting (vision -> time series)", "notes": "Cross-modal TSFM; SOTA gains of 6-44% over specialized TSFMs." }, { "name": "Moirai 2.0", "family": "Moirai", "org": "Salesforce AI Research", "release_date": "2025-08-15", "paper_arxiv_id": "2511.11698", "paper_url": "https://arxiv.org/abs/2511.11698", "params": "~10M (R-small)", "weights": "open-research-only", "architecture": "decoder-only transformer with quantile loss + multi-token prediction", "task_focus": "universal forecasting", "notes": "Switched from encoder to decoder-only; 2x faster, 30x smaller than Moirai 1.0-Large; #1 by MASE on GIFT-Eval (non-leaking)." }, { "name": "TimesFM 2.5", "family": "TimesFM", "org": "Google Research", "release_date": "2025-09-15", "paper_arxiv_id": "2310.10688", "paper_url": "https://huggingface.co/google/timesfm-2.5-200m-pytorch", "params": "200M", "weights": "open", "architecture": "decoder-only transformer with patched input + native probabilistic output", "task_focus": "general forecasting", "notes": "Halved params (500M -> 200M), 8x context length (16K), led GIFT-Eval at release on both MASE and CRPS." }, { "name": "Aurora", "family": null, "org": "East China Normal University (ECNU)", "release_date": "2025-09-26", "paper_arxiv_id": "2509.22295", "paper_url": "https://arxiv.org/abs/2509.22295", "params": "unknown", "weights": "open", "architecture": "modality-guided multi-head self-attention with prototype-guided flow matching", "task_focus": "multimodal forecasting (text + image + time series)", "notes": "ICLR 2026; first pretrained multimodal TSFM; 27-31% MSE reduction on TimeMMD vs Sundial/VisionTS." }, { "name": "CoRA (covariate adapter)", "family": null, "org": "Tsinghua University", "release_date": "2025-10-14", "paper_arxiv_id": "2510.12681", "paper_url": "https://arxiv.org/abs/2510.12681", "params": "lightweight adapter on top of existing TSFMs", "weights": "open", "architecture": "covariate-aware adapter with Granger causality embedding", "task_focus": "exogenous-covariate adaptation of any TSFM", "notes": "Plug-in adapter that lets TSFMs (Chronos, TimesFM, etc.) ingest text/image/TS covariates; 31% MSE reduction." }, { "name": "Chronos-2", "family": "Chronos", "org": "Amazon (AWS AI Labs)", "release_date": "2025-10-17", "paper_arxiv_id": "2510.15821", "paper_url": "https://arxiv.org/abs/2510.15821", "params": "120M (Small: 28M)", "weights": "open", "architecture": "encoder-only T5-inspired with group attention", "task_focus": "univariate, multivariate, and covariate-informed forecasting", "notes": "Released Oct 20 2025; SOTA on fev-bench, GIFT-Eval, Chronos Benchmark II; >90% win rate on covariate-informed tasks." }, { "name": "Xihe", "family": null, "org": "Ant Group / Tongji / collaborators", "release_date": "2025-10-20", "paper_arxiv_id": "2510.21795", "paper_url": "https://arxiv.org/abs/2510.21795", "params": "9.5M / 94M / 1.5B", "weights": "open", "architecture": "Hierarchical Interleaved Block Attention (HIBA) - multi-scale sparse attention", "task_focus": "general zero-shot forecasting", "notes": "First multi-scale TSFM; Xihe-max (1.5B) is SOTA zero-shot on GIFT-Eval; tiny/lite efficient variants." }, { "name": "TimeGPT-2", "family": "TimeGPT", "org": "Nixtla", "release_date": "2025-10-14", "paper_arxiv_id": null, "paper_url": "https://www.nixtla.io/blog/timegpt-2-announcement", "params": "unknown (Mini / standard / Pro)", "weights": "closed-api", "architecture": "transformer (architecture not disclosed)", "task_focus": "enterprise forecasting (privacy-first, on-prem)", "notes": "Private preview announced 14 Oct 2025; up to 60% accuracy improvement vs TimeGPT-1; on-prem self-hosting." }, { "name": "Moirai 2.0 paper (arxiv)", "family": "Moirai", "org": "Salesforce AI Research", "release_date": "2025-11-14", "paper_arxiv_id": "2511.11698", "paper_url": "https://arxiv.org/abs/2511.11698", "params": "~10M", "weights": "open-research-only", "architecture": "decoder-only transformer", "task_focus": "universal forecasting", "notes": "Detailed technical paper for Moirai 2.0 (HF release was Aug 2025)." }, { "name": "TimeGPT 2.1", "family": "TimeGPT", "org": "Nixtla", "release_date": "2025-12-15", "paper_arxiv_id": null, "paper_url": "https://www.nixtla.io/blog/timegpt-2-1-announcement", "params": "unknown", "weights": "closed-api", "architecture": "transformer (multivariate)", "task_focus": "enterprise multivariate forecasting", "notes": "First multivariate model in the TimeGPT family." }, { "name": "Cisco Time Series Model 1.0-preview", "family": "Cisco TSM", "org": "Cisco / Splunk", "release_date": "2025-11-25", "paper_arxiv_id": "2511.19841", "paper_url": "https://arxiv.org/abs/2511.19841", "params": "unknown", "weights": "open", "architecture": "decoder-only transformer with multi-resolution patches (TimesFM-style backbone)", "task_focus": "observability & security forecasting", "notes": "First open-weights Cisco TSFM; trained on 300B+ data points (>50% proprietary observability)." }, { "name": "Timer-S1", "family": "Timer", "org": "Tsinghua University (THUML)", "release_date": "2026-03-05", "paper_arxiv_id": "2603.04791", "paper_url": "https://arxiv.org/abs/2603.04791", "params": "8.3B (0.75B activated)", "weights": "open", "architecture": "MoE decoder-only transformer with TimeMoE + TimeSTP blocks (Serial Scaling)", "task_focus": "general forecasting", "notes": "Billion-scale TSFM with 11.5K context length; trained on TimeBench (1T points)." }, { "name": "Toto 2.0", "family": "Toto", "org": "Datadog", "release_date": "2026-04-15", "paper_arxiv_id": null, "paper_url": "https://www.datadoghq.com/blog/ai/toto-2/", "params": "4M / 22M / 313M / 1B / 2.5B", "weights": "open", "architecture": "decoder-only transformer (u-muP scaling family)", "task_focus": "observability + general forecasting", "notes": "Public blog/launch May 14 2026; HF release April 2026; SOTA on BOOM, GIFT-Eval, and TIME benchmarks; 7x more parameter-efficient than Toto 1.0." } ]