--- title: TimeGPT alternatives in 2026: hosted and open-source forecasting models description: TimeGPT compared with TimesFM, Chronos-2, Toto 2, Moirai, TiRex-2, Granite TTM, statsforecast, Darts and AutoGluon: hosting, licence, covariates. date: 2026-10-07 updated: 2026-10-07 summary: TimeGPT is a closed, hosted forecasting model from Nixtla; most alternatives are open-weights models you can run yourself or reach through a cloud service. Google's TimesFM runs inside BigQuery, Amazon's Chronos-2 deploys through SageMaker JumpStart or AutoGluon-Cloud, and Toto 2, TiRex-2 and Granite models are open weights with no first-party API we could find. Licences differ: TimesFM-3 and Moirai 2.0 weights are non-commercial. tags: forecasting, timegpt, alternatives, open-source draft: false --- ## Key facts - Third-party facts checked on 2026-10-07. Each links to the source we opened in the body below; cells we could not verify say "unknown". - TimeGPT's weights are not public. Nixtla serves it as a hosted API and offers a self-hosted version (Docker or Python wheel) on request. - Nixtla's docs list TimeGPT-2 family model IDs (`timegpt-2-pro`, `timegpt-2-lab`, `timegpt-2-mini`, `timegpt-2.1`) and ask users to confirm access with Nixtla support first. - TimesFM 2.5 is GA in BigQuery's `AI.FORECAST`; TimesFM 3.0 with covariates is in Preview there. TimesFM 3.0's open weights are under a non-commercial licence. - Chronos-2 (Apache-2.0, 120M parameters) deploys on AWS through SageMaker JumpStart or AutoGluon-Cloud. - Moirai 2.0 weights are CC-BY-NC-4.0 and, per its paper, drop multivariate and covariate support. - Toto 2 and TiRex-2 are Apache-2.0 open weights; only TiRex-2 takes covariates. ## What is TimeGPT, and why look for an alternative? TimeGPT is Nixtla's pretrained forecasting model, served over an API: you send history, it returns forecasts with prediction intervals. It is closed: the weights are not published, though the Python SDK is Apache-2.0 on [GitHub](https://github.com/Nixtla/nixtla). Per the [Nixtla FAQ](https://www.nixtla.io/docs/introduction/faq), TimeGPT supports prediction intervals and quantiles, takes exogenous variables (Nixtla's term for covariates: extra series such as price or holidays that help explain the target), and comes in a self-hosted version via Docker or a Python wheel if you contact Nixtla. The [TimeGPT-2 docs](https://www.nixtla.io/docs/forecasting/timegpt_2_family) list four model IDs and say to "confirm with support@nixtla.io that your account has access to these latest models". Nixtla [announced TimeGPT-2.1](https://www.nixtla.io/blog/timegpt-2-1-announcement) as the first multivariate model in the family. Sources disagree on whether 2.1 is generally available, so check with Nixtla. We found no public price list. The FAQ mentions free credits and says to contact Nixtla for a paid plan. TimeGPT-1 is also listed in the [Microsoft Foundry catalog](https://ai.azure.com/catalog/models/TimeGPT-1) as generally available, with the note "Legacy model. Intended for testing or benchmarking purposes." Nixtla has [announced MCP tools for Nixtla Enterprise](https://www.nixtla.io/blog/genai-announcement) (December 2025) that work with Claude Desktop, Claude Code and Cursor. Reasons people look elsewhere: wanting open weights they can run and audit, wanting a published price, or wanting a model from a different publisher to compare against. TimeGPT remains a reasonable choice if you want a single vendor with support, exogenous variables and an on-premises option. ## How do the alternatives compare? The table covers the questions that usually decide it. "Hosted" means a first-party service runs it for you. "Agent access" means an official MCP server or agent tool from the publisher. | Option | Hosted | Open weights | Licence | Probabilistic | Covariates | MCP or agent access | |---|---|---|---|---|---|---| | Nixtla TimeGPT | Yes (Nixtla API; TimeGPT-1 on Microsoft Foundry) | No | Proprietary; SDK Apache-2.0 | Yes | Yes | MCP tools announced for Nixtla Enterprise | | Google TimesFM 2.5 | Yes (BigQuery AI.FORECAST, GA) | Yes | Apache-2.0 | Yes | Library only (XReg) | BigQuery Forecast tool in MCP Toolbox | | Google TimesFM 3.0 | Preview in BigQuery | Yes | Non-commercial (weights) | Yes | Yes (BigQuery Preview) | unknown | | Amazon Chronos-2 | Deploy in your AWS account (SageMaker JumpStart, AutoGluon-Cloud) | Yes | Apache-2.0 | Yes | Yes | unknown | | Datadog Toto 2 | unknown | Yes | Apache-2.0 | Yes | No (planned) | unknown | | Salesforce Moirai 2.0 | unknown | Yes | CC-BY-NC-4.0 (small checkpoint) | Yes | No | unknown | | NXAI TiRex-2 | unknown | Yes | Apache-2.0 | Yes | Yes | unknown | | IBM Granite TTM r2 | unknown | Yes | Apache-2.0 | No (point forecasts) | With fine-tuning | unknown | | statsforecast (library) | No | Not applicable | Apache-2.0 | Yes | Yes (some models) | unknown | | Darts (library) | No | Not applicable | Apache-2.0 | Yes | Yes | unknown | | AutoGluon-TimeSeries (library) | No | Not applicable | Apache-2.0 | Yes | Yes | unknown | | Ephemeris | Yes (REST API) | Serves open-weights models | Served models Apache-2.0 or OpenMDW-1.0 | Yes | Yes (Chronos-2, TiRex-2) | Yes (remote MCP server) | "Probabilistic" means the model returns quantiles or prediction intervals, not just one line. A quantile is a level the future value should fall below with a given probability; the 0.1 and 0.9 quantiles together give an 80% band. ## Can I get Google TimesFM as a hosted service? Yes, inside BigQuery, which makes TimesFM the closest alternative if your data already lives there. Google's [November 2025 post](https://cloud.google.com/blog/products/data-analytics/timesfm-models-in-bigquery-and-alloydb) says `AI.FORECAST` is GA in BigQuery with TimesFM 2.5, with context windows from 64 to 15K points. TimesFM 3.0 is available in BigQuery as a Preview under Pre-GA terms, with `past_covariate_cols` and `future_covariate_cols`, per Google's [multivariate tutorial](https://docs.cloud.google.com/bigquery/docs/timesfm-multivariate-single-time-series-forecasting-tutorial). Google also offers a [BigQuery Forecast tool in the MCP Toolbox](https://cloud.google.com/blog/products/data-analytics/ai-based-forecasting-and-analytics-in-bigquery-via-mcp-and-adk) so an agent can run `AI.FORECAST`. The weights are on Hugging Face. [TimesFM 2.5](https://huggingface.co/google/timesfm-2.5-200m-pytorch) is Apache-2.0 at 200M parameters. [TimesFM 3.0](https://huggingface.co/google/timesfm-3.0-pytorch) is 0.3B parameters under the TimesFM Non-Commercial License v1.0. The [TimesFM README](https://github.com/google-research/timesfm) says commercial and production use of 3.0 is permitted through Google Cloud services such as BigQuery ML. TimesFM-3 also has the best GIFT-Eval score in the table on [our benchmarks page](/benchmarks): CRPS 0.4557 on the published GIFT-Eval leaderboard (97 configurations, ratio to seasonal naive, lower is better). In our view, if you are on Google Cloud and can accept Preview terms, it is the first thing to try. ## How does AWS offer Chronos-2? As open weights you deploy yourself, not as a per-call API. Chronos-2 is a 120M-parameter Apache-2.0 model that handles univariate, multivariate and covariate-informed forecasting, including categorical covariates, per its [model card](https://huggingface.co/amazon/chronos-2). The card reports over 300 forecasts per second on a single A10G GPU. The [Chronos repository](https://github.com/amazon-science/chronos-forecasting) recommends AutoGluon-Cloud (real-time, serverless and batch inference) or SageMaker JumpStart endpoints on CPU or GPU, both running in your own AWS account. The older Chronos-Bolt is on [Amazon Bedrock Marketplace](https://aws.amazon.com/blogs/machine-learning/how-deutsche-bahn-redefines-forecasting-using-chronos-models-now-available-on-amazon-bedrock-marketplace/); we could not confirm Chronos-2 there. Choose this route if you are on AWS and want the endpoint in your own account. ## Is Datadog Toto 2 available as an API? We found no hosted API from Datadog: Toto 2.0 is open weights, [released on 14 May 2026](https://www.datadoghq.com/blog/ai/toto-2/) in five sizes from 4M to 2.5B parameters, Apache-2.0. It is multivariate and returns nine quantiles, per the [313M model card](https://huggingface.co/Datadog/Toto-2.0-313m). The [Toto repository](https://github.com/DataDog/toto) says exogenous variable support for 2.0 is "planned for a future 2.0 release but not yet available"; Toto 1.0 (151M) has it. ## Can I use Salesforce Moirai 2.0 commercially? Not under its licence. [Moirai 2.0 small](https://huggingface.co/Salesforce/moirai-2.0-R-small) is an 11.4M-parameter model under CC-BY-NC-4.0, a non-commercial licence. The [Moirai 2.0 paper](https://arxiv.org/html/2511.11698v1) says Salesforce "dropped support for multivariate forecasting and the use of covariates in Moirai 2.0". The training code, [uni2ts](https://github.com/SalesforceAIResearch/uni2ts), is Apache-2.0. It is a fit for research and benchmarking, not for a commercial product under that licence. ## What is NXAI TiRex-2 good for? Covariate-aware forecasting in a small model. [TiRex-2](https://huggingface.co/NX-AI/TiRex-2) is Apache-2.0, 38.4M parameters univariate plus 44.1M for multivariate, and conditions on past and future-known covariates. Its predecessor TiRex is under the NXAI community licence, per the [TiRex repository](https://github.com/NX-AI/tirex), so check which one you use. NXAI's [TiRex-2 page](https://www.nx-ai.com/en/tirex-2) describes the open model as "free to use and explore, but not customizable", with a Pro tier for fine-tuning, streaming and edge variants. We found no public hosted API. ## What are IBM Granite TTM models? [Granite TTM r2](https://huggingface.co/ibm-granite/granite-timeseries-ttm-r2) models are Apache-2.0 and start at 805K parameters; the card says they run on a laptop or a single GPU. They produce point forecasts, not quantiles, and use exogenous variables when fine-tuned. IBM Granite also publishes PatchTST-FM r2 and FlowState r1, both probabilistic. We cover them in [the decision guide](/blog/which-forecasting-model-should-i-use). ## Should I use a library instead of a foundation model? Often, yes. In our experience, if you have many series with long, regular histories, classical statistical models are fast, cheap and a baseline any foundation model should have to beat. - [statsforecast](https://github.com/Nixtla/statsforecast) (Nixtla, Apache-2.0): AutoARIMA, Theta and other statistical models, with prediction intervals and exogenous variables on several models. - [Darts](https://github.com/unit8co/darts) (Apache-2.0): one interface over statistical, machine-learning and foundation models, with past, future and static covariates and probabilistic forecasts. It wraps Chronos-2, TimesFM 2.5 and 3, TiRex and PatchTST-FM. - [AutoGluon-TimeSeries](https://auto.gluon.ai/stable/tutorials/timeseries/forecasting-model-zoo.html) (Apache-2.0, per the [AutoGluon repository](https://github.com/autogluon/autogluon)): automatic model selection and ensembling, including Chronos-2 with covariates, Chronos, Toto and Toto 2.0. Libraries are the better choice when you want everything on your own hardware, when you will fine-tune, or when you want to benchmark foundation models against classical baselines on your own data. ## When is self-hosting an open-weights model the better choice? Self-host when your data cannot leave your network, when you forecast at a volume where a GPU you already pay for is cheaper than per-call pricing, or when you need to fine-tune. All the open-weights models above are on Hugging Face. The cost is operational: you run the GPU, pin library versions, and track new releases. The Granite TTM card, for example, says "IBM is under no obligation to provide enhancements, updates, or support", so you own that maintenance. ## Where does Ephemeris fit? Ephemeris is a hosted API and remote MCP server over six open-weights models: Chronos-2, TimesFM 2.5, Toto 2, TiRex-2, PatchTST-FM r2 and FlowState r1. You can name one model, let it route, or ask for an accuracy-weighted ensemble. Pricing is prepaid credits per model run, published on [/pricing](/pricing). It is not the right choice if you need TimesFM-3 (not served), fine-tuning, or on-premises deployment. Our benchmark numbers are our own runs with each benchmark's harness, not leaderboard submissions; we have not benchmarked TimeGPT. ## FAQ ### Is TimeGPT open source? No. TimeGPT's weights are not published; Nixtla's Python SDK is Apache-2.0. Nixtla offers a self-hosted version via Docker or a Python wheel on request. ### Which TimeGPT alternatives are free for commercial use? Among open-weights models, Chronos-2, TimesFM 2.5, Toto 2 and TiRex-2 are all Apache-2.0. TimesFM 3.0 and Moirai 2.0 have open weights but non-commercial licences. ### Can I use TimesFM commercially? TimesFM 2.5 weights are Apache-2.0. TimesFM 3.0 weights are non-commercial, but Google's README says commercial use of 3.0 is permitted through Google Cloud services such as BigQuery ML. ### Which TimeGPT alternatives support covariates? Chronos-2 and TiRex-2 accept past and future-known covariates out of the box. TimesFM supports them through XReg in its library and through TimesFM 3.0 in BigQuery (Preview); Toto 2 and Moirai 2.0 do not. ### Does Amazon offer Chronos as an API? Not as a per-call API. Chronos-2 deploys into your own AWS account through SageMaker JumpStart or AutoGluon-Cloud; Chronos-Bolt is on Amazon Bedrock Marketplace. ## Related - [Which time-series foundation model should I use?](/blog/which-forecasting-model-should-i-use) - [Chronos-2 vs TimesFM 2.5 vs Toto 2 vs TiRex-2](/blog/chronos-2-vs-timesfm-vs-toto-vs-tirex) - [A guide to time-series foundation models](/blog/time-series-foundation-models-guide) - [A forecasting tool for AI agents over MCP](/blog/forecasting-tool-for-ai-agents-mcp) - [Model pages](/models) and [pricing](/pricing)