Catalog

117 models a TServe process can load. The server always loads naive, a no-download baseline for testing. Name a catalog model as a leftover positional for a real forecast. GET /models reports what did. The same model goes in leftover CLI positionals and in a request's model field.

Dependencies

The extra name is the CPU image tag. GPU tags are {extra}-gpu. There is no :base-gpu.

kronos is built on base, so it cannot load Chronos Bolt, TTM, or TimesFM. hub can, and so can every extra that includes it: chronos, granite, moirai, tirex, tirex2, toto, mantis, timesfm3, t0, tafsut, and full.

full is chronos, kronos, granite, moirai, tirex, tirex2, toto, mantis, timesfm3, t0, and tafsut. client, http, dev, docs, and all-extras are not model families. Moirai pins gluonts, lightning, and hydra-core when python_version < '3.14'.

added counts checkpoints that extra contributes. full is the total, including naive.

extra	CPU tag	GPU tag	families	added	example
server	:base	—	Naive	1	naive
hub	:hub	:hub-gpu	Chronos Bolt, Chronos T5, TTM, TimesFM 2.x	81	chronos_bolt
chronos	:chronos	:chronos-gpu	Chronos-2	3	chronos_2
kronos	:kronos	:kronos-gpu	Kronos, WindFM	5	kronos
granite	:granite	:granite-gpu	FlowState	2	flowstate
moirai	:moirai	:moirai-gpu	Moirai 2, Moirai 1.x, Lag-Llama	8	moirai_2
tirex	:tirex	:tirex-gpu	TiRex	2	tirex
tirex2	:tirex2	:tirex2-gpu	TiRex-2	4	tirex_2
toto	:toto	:toto-gpu	Toto-2	5	toto_2_0_4m
mantis	:mantis	:mantis-gpu	Mantis	3	mantis_8m
timesfm3	:timesfm3	:timesfm3-gpu	TimesFM 3	1	timesfm_3
t0	:t0	:t0-gpu	T0	1	t0
tafsut	:tafsut	:tafsut-gpu	Tafsut	1	tafsut
full	:full	:full-gpu	all of the above	117	chronos_2

Start a server

basehubchronoskronosgranitemoiraitirextotomantistimesfm3tirex2t0tafsutfull

naive only — a test baseline, nothing downloaded. Full page: base.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:base

uv pip install "tserve[server]"

uv run tserve

pip install "tserve[server]"

tserve

Chronos Bolt, Chronos T5, TTM, TimesFM 2.x: 81 models, plus naive. Full page: hub.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:hub chronos_bolt

uv pip install "tserve[server,hub]"

uv run tserve chronos_bolt

pip install "tserve[server,hub]"

tserve chronos_bolt

Chronos-2, plus every hub model. Full page: chronos.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:chronos chronos_2

uv pip install "tserve[server,chronos]"

uv run tserve chronos_2

pip install "tserve[server,chronos]"

tserve chronos_2

Kronos and WindFM. Sits on base, so no Hub models. Full page: kronos.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:kronos kronos

uv pip install "tserve[server,kronos]"

uv run tserve kronos

pip install "tserve[server,kronos]"

tserve kronos

FlowState, plus every hub model. Full page: granite.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:granite flowstate

uv pip install "tserve[server,granite]"

uv run tserve flowstate

pip install "tserve[server,granite]"

tserve flowstate

Moirai 2, Moirai 1.x, Lag-Llama, plus every hub model. Full page: moirai.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:moirai moirai_2

uv pip install "tserve[server,moirai]"

uv run tserve moirai_2

pip install "tserve[server,moirai]"

tserve moirai_2

TiRex, plus every hub model. Full page: tirex.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:tirex tirex

uv pip install "tserve[server,tirex]"

uv run tserve tirex

pip install "tserve[server,tirex]"

tserve tirex

Toto-2, plus every hub model. Full page: toto.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:toto toto_2_0_4m

uv pip install "tserve[server,toto]"

uv run tserve toto_2_0_4m

pip install "tserve[server,toto]"

tserve toto_2_0_4m

Mantis, plus every hub model. Needs past longer than 127 rows. Full page: mantis.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:mantis mantis_8m

uv pip install "tserve[server,mantis]"

uv run tserve mantis_8m

pip install "tserve[server,mantis]"

tserve mantis_8m

TimesFM 3, plus every hub model. Full page: timesfm3.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:timesfm3 timesfm_3

uv pip install "tserve[server,timesfm3]"

uv run tserve timesfm_3

pip install "tserve[server,timesfm3]"

tserve timesfm_3

TiRex-2, plus every hub model. Full page: tirex2.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:tirex2 tirex_2

uv pip install "tserve[server,tirex2]"

uv run tserve tirex_2

pip install "tserve[server,tirex2]"

tserve tirex_2

T0, plus every hub model. The checkpoint is gated. Full page: t0.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:t0 t0

uv pip install "tserve[server,t0]"

uv run tserve t0

pip install "tserve[server,t0]"

tserve t0

Tafsut, plus every hub model. Full page: tafsut.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:tafsut tafsut

uv pip install "tserve[server,tafsut]"

uv run tserve tafsut

pip install "tserve[server,tafsut]"

tserve tafsut

All 117 models, and the only way to mix two family stacks. Full page: full.

Dockeruvpip

docker run --rm -p 8000:8000 sktime/tserve:full \
  chronos_2 tirex kronos

uv pip install "tserve[server,full]"

uv run tserve chronos_2 tirex kronos

pip install "tserve[server,full]"

tserve chronos_2 tirex kronos

• Docker tags, token, and cache volume: Docker. Each CPU tag above has a -gpu variant for NVIDIA hosts (GPU images).
• uv and pip install from PyPI: UV / Pip. From a clone: From source.
• Ids not listed here: craft specs, live objects, a directory of .zip files.

Predict

bash / zshPowerShellPython

curl -s http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d '{
  "past": {
    "timestamp": ["2024-01-01", "2024-01-02", "2024-01-03", "2024-01-04", "2024-01-05"],
    "sales": [120, 135, 128, 142, 138]
  },
  "time": "timestamp",
  "target": ["sales"],
  "fh": 3,
  "model": "chronos_bolt"
}'

curl.exe -s http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d '{"past":{"timestamp":["2024-01-01","2024-01-02","2024-01-03","2024-01-04","2024-01-05"],"sales":[120,135,128,142,138]},"time":"timestamp","target":["sales"],"fh":3,"model":"chronos_bolt"}'

from tserve.client import Client

past = {
    "timestamp": ["2024-01-01", "2024-01-02", "2024-01-03", "2024-01-04", "2024-01-05"],
    "sales": [120, 135, 128, 142, 138],
}

with Client("http://127.0.0.1:8000") as client:
    result = client.predict(
        past=past,
        time="timestamp",
        target=["sales"],
        fh=3,
        model="chronos_bolt",
    )
print(result.predictions)

Only model changes between extras. Python needs the client extra on the caller. Every field, format, and response shape: Data specification.

All models

Naive

NaiveForecaster extra server 1 model

multivariate exogenous quantiles

Drift strategy. No weights, no Hugging Face download. Always loaded so you can test the server; name another catalog model for a real forecast.

model	checkpoint
naive	none

Chronos Bolt

ChronosForecaster extra hub 4 models

multivariate exogenous quantiles

model	checkpoint
chronos_bolt	amazon/chronos-bolt-tiny
chronos_bolt_mini	amazon/chronos-bolt-mini
chronos_bolt_small	amazon/chronos-bolt-small
chronos_bolt_base	amazon/chronos-bolt-base

Chronos T5

ChronosForecaster extra hub 5 models

multivariate exogenous quantiles

The original Chronos line.

model	checkpoint
chronos_t5	amazon/chronos-t5-tiny
chronos_t5_mini	amazon/chronos-t5-mini
chronos_t5_small	amazon/chronos-t5-small
chronos_t5_base	amazon/chronos-t5-base
chronos_t5_large	amazon/chronos-t5-large

TTM

TinyTimeMixerForecaster extra hub 70 models

multivariate exogenous quantiles

IBM Granite Tiny Time Mixers. Models are {revision}-{context}-{horizon}, with optional -lite or -l1. The four short models instead take the forecaster default revision, and ttm its default repo too.

TTM defaults

model	repo	revision
ttm	forecaster default (ibm/TTM)	forecaster default (main)
ttm_r1	ibm-granite/granite-timeseries-ttm-r1	forecaster default (main)
ttm_r2	ibm-granite/granite-timeseries-ttm-r2	forecaster default (main)
ttm_r3	ibm-granite/granite-timeseries-ttm-r3	forecaster default (main)

TTM r1

ibm-granite/granite-timeseries-ttm-r1

model	context	horizon
ttm_r1_512_96	512	96
ttm_r1_1024_96	1024	96

TTM r2

ibm-granite/granite-timeseries-ttm-r2

model	context	horizon
ttm_r2_512_96	512	96
ttm_r2_512_192	512	192
ttm_r2_512_336	512	336
ttm_r2_512_720	512	720
ttm_r2_1024_96	1024	96
ttm_r2_1024_192	1024	192
ttm_r2_1024_336	1024	336
ttm_r2_1024_720	1024	720
ttm_r2_1536_96	1536	96
ttm_r2_1536_192	1536	192
ttm_r2_1536_336	1536	336
ttm_r2_1536_720	1536	720

TTM r2.1

Same Hub repo as r2. -l1 is the L1 checkpoint.

model	context	horizon	variant
ttm_r2_1_52_16	52	16
ttm_r2_1_52_16_l1	52	16	L1
ttm_r2_1_90_30	90	30
ttm_r2_1_90_30_l1	90	30	L1
ttm_r2_1_180_60_l1	180	60	L1
ttm_r2_1_360_60_l1	360	60	L1
ttm_r2_1_512_48	512	48
ttm_r2_1_512_48_l1	512	48	L1
ttm_r2_1_512_96	512	96
ttm_r2_1_512_96_l1	512	96	L1

TTM r3

ibm-granite/granite-timeseries-ttm-r3. Each model has a -lite sibling.

model	lite	context	horizon
ttm_r3_52_16	ttm_r3_52_16_lite	52	16
ttm_r3_90_30	ttm_r3_90_30_lite	90	30
ttm_r3_156_16	ttm_r3_156_16_lite	156	16
ttm_r3_180_60	ttm_r3_180_60_lite	180	60
ttm_r3_360_60	ttm_r3_360_60_lite	360	60
ttm_r3_512_30	ttm_r3_512_30_lite	512	30
ttm_r3_512_48	ttm_r3_512_48_lite	512	48
ttm_r3_512_96	ttm_r3_512_96_lite	512	96
ttm_r3_512_336	ttm_r3_512_336_lite	512	336
ttm_r3_768_48	ttm_r3_768_48_lite	768	48
ttm_r3_1024_48	ttm_r3_1024_48_lite	1024	48
ttm_r3_1024_96	ttm_r3_1024_96_lite	1024	96
ttm_r3_1024_720	ttm_r3_1024_720_lite	1024	720
ttm_r3_1536_96	ttm_r3_1536_96_lite	1536	96
ttm_r3_1536_720	ttm_r3_1536_720_lite	1536	720
ttm_r3_2048_96	ttm_r3_2048_96_lite	2048	96
ttm_r3_2048_720	ttm_r3_2048_720_lite	2048	720
ttm_r3_2560_96	ttm_r3_2560_96_lite	2560	96
ttm_r3_2560_720	ttm_r3_2560_720_lite	2560	720
ttm_r3_3072_96	ttm_r3_3072_96_lite	3072	96
ttm_r3_3072_720	ttm_r3_3072_720_lite	3072	720

TimesFM 2.x

TimesFM2Forecaster extra hub 2 models

multivariate exogenous quantiles

Quantile levels are the ones the checkpoint config ships.

model	checkpoint
timesfm_2_5	google/timesfm-2.5-200m-transformers
timesfm_2	google/timesfm-2.0-500m-pytorch

TimesFM 3

TimesFM3Forecaster extra timesfm3 1 model

multivariate exogenous quantiles

The registry sets license_accepted=True. Weights use the TimesFM non-commercial license. Point forecasts are the median. Native quantile levels are 0.1 through 0.9.

model	checkpoint
timesfm_3	google/timesfm-3.0-pytorch

T0

T0Forecaster extra t0 1 model

multivariate exogenous quantiles

The registry sets license_accepted=True. The checkpoint is gated. Each target is forecast on its own. Point forecasts are the median.

model	checkpoint
t0	theforecastingcompany/t0-alpha

Tafsut

TafsutForecaster extra tafsut 1 model

multivariate exogenous quantiles

Native quantile levels are 0.1 through 0.9.

model	checkpoint
tafsut	Tafsut-FM/tafsut-univariate-base

Chronos-2

Chronos2Forecaster extra chronos 3 models

multivariate exogenous quantiles

model	checkpoint
chronos_2	amazon/chronos-2
chronos_2_small	autogluon/chronos-2-small
chronos_2_synth	autogluon/chronos-2-synth

Kronos

KronosForecaster extra kronos 3 models

multivariate exogenous quantiles

The tokenizer ships with the model, not as a separate model.

model	checkpoint
kronos	NeoQuasar/Kronos-small
kronos_mini	NeoQuasar/Kronos-mini
kronos_base	NeoQuasar/Kronos-base

WindFM

WindFMForecaster extra kronos 2 models

multivariate exogenous quantiles

model	checkpoint
windfm	NeoQuasar/WindFM
windfm_robust	NeoQuasar/WindFM-robust

FlowState

FlowStateForecaster extra granite 2 models

multivariate exogenous quantiles

Revision pinned to r1.1.

model	checkpoint
flowstate	ibm-research/flowstate
flowstate_granite	ibm-granite/granite-timeseries-flowstate-r1

Moirai 2

Moirai2Forecaster extra moirai 1 model

multivariate exogenous quantiles

model	checkpoint
moirai_2	Salesforce/moirai-2.0-R-small

Moirai 1.x

MOIRAIForecaster extra moirai 6 models

multivariate exogenous quantiles

model	checkpoint
moirai_1_0_r_small	Salesforce/moirai-1.0-R-small
moirai_1_0_r_base	Salesforce/moirai-1.0-R-base
moirai_1_0_r_large	Salesforce/moirai-1.0-R-large
moirai_1_1_r_small	Salesforce/moirai-1.1-R-small
moirai_1_1_r_base	Salesforce/moirai-1.1-R-base
moirai_1_1_r_large	Salesforce/moirai-1.1-R-large

Lag-Llama

LagLlamaForecaster extra moirai 1 model

multivariate exogenous quantiles

model	checkpoint
lagllama	time-series-foundation-models/Lag-Llama

TiRex

TiRexForecaster extra tirex 2 models

multivariate exogenous quantiles

The registry sets license_accepted=True.

model	checkpoint
tirex	NX-AI/TiRex
tirex_1_1_gifteval	NX-AI/TiRex-1.1-gifteval

TiRex-2

TiRex2Forecaster extra tirex2 4 models

multivariate exogenous quantiles

Native quantile levels are 0.1 through 0.9. The three decontaminated checkpoints are gated on Hugging Face.

model	checkpoint
tirex_2	NX-AI/TiRex-2
tirex_2_gifteval_zs	NX-AI/TiRex-2-gifteval-zs
tirex_2_gifteval_pretrain	NX-AI/TiRex-2-gifteval-pretrain
tirex_2_fevbench	NX-AI/TiRex-2-fevbench

Toto-2

Toto2Forecaster extra toto 5 models

multivariate exogenous quantiles

model	checkpoint
toto_2_0_4m	Datadog/Toto-2.0-4m
toto_2_0_22m	Datadog/Toto-2.0-22m
toto_2_0_313m	Datadog/Toto-2.0-313m
toto_2_0_1b	Datadog/Toto-2.0-1B
toto_2_0_2_5b	Datadog/Toto-2.0-2.5B

Mantis

MantisForecaster extra mantis 3 models

multivariate exogenous quantiles

Embeddings plus an sklearn head. context_length is 127, so past must be longer than that.

model	checkpoint
mantis	paris-noah/MantisV2
mantis_8m	paris-noah/Mantis-8M
mantis_plus	paris-noah/MantisPlus

Capabilities

Every capability is read from the estimator, not set by a TServe flag. Multivariate is more than one target; exogenous is a covariate the model actually uses instead of ignoring; quantiles are prediction intervals (HTTP, Python).

family	multivariate	exogenous	quantiles
Naive
Chronos-2
Chronos Bolt
Chronos T5
TTM
TimesFM 2.x
TimesFM 3
T0
Tafsut
Kronos
WindFM
FlowState
Moirai 2
Moirai 1.x
Lag-Llama
TiRex
TiRex-2
Toto-2
Mantis

Where to next

• Run a server

  ---

  Flags, Server, server.app, and the process lifecycle.

  Server

• Docker

  ---

  Tags, Hugging Face token, cache volume, GPU, and building images.

  Docker

• Send predictions

  ---

  JSON over POST /predict, or native tables from Python.

  HTTP or Python

• Beyond the catalog

  ---

  Craft specs, live objects, and saved .zip models.

  Craft specs