Models

Chronos-2 is Amazon's universal successor to the original Chronos family. The official card describes it as a 120M-parameter, T5-inspired encoder-only model that handles univariate, multivariate, and covariate-informed forecasting in a single architecture. It is the Chronos variant to choose when you need cross-learning across related series, native covariates, or broader zero-shot coverage than the earlier univariate-first checkpoints.

512 context | $0.5000 input tokens | $1.50 output tokens

Chronos-Bolt Base is the larger high-accuracy checkpoint in Amazon's Chronos-Bolt family. The official model card positions it as more accurate than the original Chronos-Large while remaining dramatically faster thanks to patch-based direct forecasting. It is the better Chronos choice when you want a stronger quality ceiling without moving to the older autoregressive Chronos-T5 stack.

512 context | $0.5000 input tokens | $1.50 output tokens

Chronos-Bolt Small is Amazon's fast zero-shot forecasting checkpoint for production use. It replaces autoregressive token-by-token rollout with direct multi-step quantile prediction over patchified history, which makes it materially faster and lighter than the original Chronos line. It is a strong default when you want probabilistic forecasts with low latency and a simpler deployment footprint.

512 context | $0.5000 input tokens | $1.50 output tokens

Chronos-T5 Large is the original large Chronos checkpoint. It scales and quantizes continuous values into tokens, then uses a T5 encoder-decoder to autoregress over future token sequences and recover probabilistic forecasts. It remains a useful reference model for the first Chronos generation, especially when you want the classic tokenized Chronos setup rather than the newer Bolt or Chronos-2 designs.

512 context | $0.5000 input tokens | $1.50 output tokens

FlowState is IBM's sampling-rate-invariant TSFM for zero-shot forecasting. Its SSM encoder and functional basis decoder let it adapt context length, target length, and sampling rate at inference time instead of being tied to a single fixed timescale. It is the right IBM model when your data arrives at inconsistent cadences or when you need one model to generalize across multiple temporal resolutions.

512 context | $0.5000 input tokens | $1.50 output tokens

This catalog entry maps the live `ttm-v1` deployment to IBM's official TTM-R1 family surface. TinyTimeMixer is a compact forecasting architecture built for fast zero-shot and few-shot forecasting on standard public benchmarks, with checkpoint specializations for specific context and prediction lengths rather than one universal dense model. It is the smallest and most deployment-friendly IBM checkpoint in the hosted catalog.

512 context | $0.5000 input tokens | $1.50 output tokens

TTM-R2 is IBM's larger-data continuation of the TinyTimeMixer line. IBM positions it as a better-performing follow-on to R1 while preserving the small, fast deployment profile that makes TinyTimeMixer practical on CPUs and lightweight hosted inference. It remains a focused family of context- and horizon-specific checkpoints rather than a single universal TSFM.

512 context | $0.5000 input tokens | $1.50 output tokens

Kairos-10M is the smallest public checkpoint in the Kairos family. The official model card and project page describe adaptive tokenization plus instance-adaptive rotary position encoding to better handle heterogeneous time-series structure and varying local information density. It is the lightest-weight way to access the Kairos design while staying in the zero-shot forecasting regime.

512 context | $0.5000 input tokens | $1.50 output tokens

Kairos-23M is the mid-size public Kairos checkpoint. It keeps the same adaptive tokenization and instance-specific positional encoding strategy as the rest of the family, while offering more capacity than the 10M model for broader zero-shot coverage. It is a good middle ground when you want the Kairos design without jumping all the way to the 50M checkpoint.

512 context | $0.5000 input tokens | $1.50 output tokens

Kairos-50M is the largest public checkpoint in the Kairos family currently hosted here. Official sources describe it as an adaptive TSFM that varies tokenization granularity and positional treatment based on the information density and temporal structure of each series. It is the Kairos checkpoint to reach for when you want the highest published capacity in that adaptive family.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.0-R-Base is the reference dense checkpoint in the original Moirai family. It keeps the any-variate masked-encoder design while scaling model capacity enough to improve general forecasting quality on heterogeneous multivariate settings. It is a good default dense Moirai checkpoint when you want stronger accuracy than the small model without jumping to the much larger large variant.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.0-R-Large is the highest-capacity dense model in the original Moirai family. It preserves the same masked-encoder, any-variate forecasting design as the smaller checkpoints, but at a scale aimed at stronger broad zero-shot performance. It is the dense Moirai option to pick when you want the first-generation architecture at its highest published capacity.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.0-R-Small is the smallest dense checkpoint in Salesforce's original Moirai family. It uses a masked-encoder forecasting architecture with any-variate attention and a probabilistic mixture output head, letting it reason over arbitrary numbers of target variables and dynamic covariates. It is the most cost-efficient way to access the first-generation Moirai design.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.1-R-Base is the mid-size update in the Moirai 1.1 family. Salesforce's official note for 1.1-R emphasizes improved low-frequency performance while keeping the familiar Moirai usage pattern. It is the cleanest option when you want Moirai's dense masked-encoder design but prefer the newer 1.1 checkpoint family over the original 1.0 release.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.1-R-Large is the largest checkpoint in the 1.1-R update line. The official card does not publish a full new technical note, but it frames 1.1-R as a direct improvement on the original 1.0-R family with better low-frequency behavior. It is the Moirai 1.1 option to choose when you want the highest-capacity checkpoint in that updated dense family.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-1.1-R-Small is the updated small checkpoint in the Moirai 1.1 line. Salesforce describes the 1.1-R release as an improvement over 1.0-R, especially for low-frequency yearly and quarterly cases on Monash-style evaluation sets. The public card is sparse, but the released configuration follows the same overall Moirai architecture and API pattern as the 1.0-R family.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-2.0-R-Small is the faster successor to the first dense Moirai family. The official Moirai 2.0 release switches to a decoder-only design with quantile loss, multi-token prediction, and better missing-value handling, while Salesforce reports performance that surpasses larger earlier Moirai checkpoints. It is the modern small-footprint Moirai option when you want the newer generation rather than the original masked-encoder line.

512 context | $0.5000 input tokens | $1.50 output tokens

Moirai-MoE-1.0-R-Base is Salesforce's sparse expert extension of the Moirai line. Instead of relying on a single dense model for all behaviors, it routes tokens through specialized experts to improve parameter efficiency and specialization across heterogeneous series. Salesforce's official Moirai-MoE materials position the Base variant as a top zero-shot performer in the family while keeping inference cheaper than an equally sized dense alternative.

512 context | $0.5000 input tokens | $1.50 output tokens

MOMENT-Large is the large checkpoint in AutonLab's general-purpose time-series foundation-model family. Official sources frame MOMENT as a multi-task representation model that transfers across forecasting, classification, anomaly detection, imputation, reconstruction, and embedding extraction rather than optimizing purely for one forecasting benchmark. It is the most flexible hosted model when you expect to reuse a shared backbone across several downstream time-series tasks.

512 context | $0.5000 input tokens | $1.50 output tokens

This checkpoint is IBM's hosted PatchTST reference model for ETTh1 rather than a broad multi-domain TSFM family card. It uses patch tokenization and Transformer blocks to model long-horizon time-series behavior efficiently, and the published checkpoint is tuned around the ETTh1 workload. It is most useful as a strong baseline and interpretable point of comparison against the larger foundation models in the catalog.

512 context | $0.5000 input tokens | $1.50 output tokens

Sundial is THUML's native generative TSFM for continuous-valued forecasting. The official card presents it as a 128M-parameter decoder-only model trained with TimeFlow Loss, built to generate multiple plausible futures without relying on discrete tokenization. It is one of the most directly probabilistic models in the hosted catalog and a strong option when distribution quality matters as much as point accuracy.

512 context | $0.5000 input tokens | $1.50 output tokens

Time-MoE-200M is a sparse Mixture-of-Experts forecasting model from the Time-MoE family. The official paper and repo position the family as a way to scale time-series model capacity more efficiently than dense decoder-only models while keeping long-context autoregressive forecasting practical. It is a good fit when you want long-context zero-shot forecasting with explicit sparse-expert scaling.

512 context | $0.5000 input tokens | $1.50 output tokens

TimesFM 2.0 500M is Google's larger open TimesFM checkpoint for zero-shot time-series forecasting. It is a decoder-only patched transformer focused primarily on univariate point forecasting, with optional experimental quantile heads that Google notes are not calibrated after pretraining. It is the stronger open TimesFM variant when raw forecast accuracy matters more than footprint.

512 context | $0.5000 input tokens | $1.50 output tokens

TimesFM 2.5 200M is Google's smaller and more deployment-friendly open TimesFM checkpoint. The official repo describes it as a 16K-context model with an optional continuous quantile head, no frequency indicator requirement, and speed-oriented structural updates such as QKV fusion. It is a strong choice when you want modern TimesFM behavior with much lower model size than the 500M release.

512 context | $0.5000 input tokens | $1.50 output tokens

TiRex is NX-AI's xLSTM-based zero-shot forecasting model. The official card highlights strong short- and long-horizon performance, plus support for both point and quantile forecasts, without relying on a Transformer backbone. It is a good alternative when you want a compact non-Transformer forecaster with strong in-context generalization behavior.

512 context | $0.5000 input tokens | $1.50 output tokens

Toto is Datadog's observability-oriented forecasting foundation model. The official model card positions it for high-dimensional, sparse, non-stationary telemetry, and reports strong results on BOOM while remaining competitive on broader benchmark suites. It is the most workload-specific model in the catalog when your data looks like production infrastructure metrics rather than clean academic series.

512 context | $0.5000 input tokens | $1.50 output tokens