TabPFN, TimesFM and friends

The idea in one line: for tables and time series, one pretrained model given your data as context now beats hours of per-task tuning on most public benchmarks, but the leader changes every few weeks and most of the best weights forbid commercial use.

The 10x decision is whether you use one of these models, not which one. Choosing an in-context model over a tuned gradient-boosted tree removes the tuning bill: about 700x less fitting time on TabArena (next section has the numbers). Choosing between the top five models is 10% polish: they sit within a few percent of each other, and the order reshuffles monthly. Google’s own charts put TimesFM-3 on top at its launch on 2026-08-31; by 2026-10-28 it was 6th among single models on GIFT-Eval.

Small is nearly as good, which makes “cheap enough to run on everything” true. Mistral’s 5M-parameter Timestral forecasts in 17 ms on a server CPU (Mistral, 2026-10-15) and still beats Google’s 200M TimesFM-2.5 on GIFT-Eval. The previous lesson explained why these models work; this one is the field guide.

Elo (TabArena) is a chess-style rating from pairwise wins: a 400-point gap means about 10:1 odds that the higher-rated method wins on a given dataset. The formula is odds = 10^(gap / 400). A 600-point gap gives 10^1.5, about 32:1.

Seasonal naive is the “do nothing clever” forecast: predict each future step as the value one season earlier. Tomorrow’s 3 pm load is today’s 3 pm load. Time-series benchmarks divide every model’s error by seasonal naive’s.

• MASE is point-forecast error divided by seasonal naive’s error; below 1 beats it.
• CRPS scores a probabilistic forecast. A model outputs quantiles (its 90th-percentile forecast is a value the outcome should fall below 90% of the time), and CRPS averages how far off each one is. Lower is better.
• GIFT-Eval reports a CRPS ratio: your CRPS over seasonal naive’s, combined across datasets with a geometric mean (multiply n ratios, take the n-th root). Ratios of 0.3 and 0.7 give sqrt(0.3 x 0.7) = 0.46, where a plain average would say 0.5. One dataset where everything blows up can’t swamp the rest.
• fev-bench reports SQL skill: 1 minus the geometric-mean error ratio. Skill 0.49 means about half of seasonal naive’s error. Higher is better; seasonal naive scores 0.

Data leakage means benchmark test data, or data highly correlated with it, sat in the pretraining set, inflating the score. GIFT-Eval now flags it per model, and fev-bench replaces a leaked task’s score with Chronos-Bolt’s, the leaderboard’s own rule.

These models take a table of labelled rows as context and predict labels or numbers for new rows. All of the leaders below are pretrained on synthetic tables only, the structural causal model recipe.

Prior Labs, founded in Germany in 2024 to commercialise TabPFN (which came out of a University of Freiburg lab), has shipped six versions.

Sources: v1, v2, 2.5, 3, 3.5, models page. The 2.5-and-later results are Prior Labs’ own; the 1M-row wins come from its internal datasets.

The architecture grew with the row limits: TabPFN-3 embeds each column, compresses each row, then runs in-context learning over rows, so 1M rows fit on one H100. In practice an 8 GB GPU suffices for most datasets, CPU use is capped at 5,000 samples, and you predict in batches of about 1,000 because each predict call rereads the training set (README).

The business moved fast. SAP agreed to buy Prior Labs on 2026-05-04, pledging over €1B across four years, 15 months after a €9M pre-seed (SAP News, 2026-05-04).

Sources: TabArena board v0.1.9.12, 2026-10-27; Kumo-Tabular; TabFM; Mitra-v2; TabICLv2.

Notes worth knowing:

• Kumo-Tabular is synthetic-only and #1. NVIDIA bought Kumo AI around June 2026, for about $400M per a single secondary report (Pulse 2.0, 2026-06-09).
• TabICLv2 is the open-science pick: 28M parameters, BSD-licensed, with its pretraining code and prior released. It fits and predicts a 50,000 x 100 table in under 10 seconds on an H100, 10.6x faster than TabPFN-2.5, and untuned it edges the tuned, real-data fine-tuned Real-TabPFN-2.5 (1591 versus 1587).
• Mitra v1 (July 2025) mixed causal-model priors with tree-based priors and ships in AutoGluon 1.4. TabFM’s size is disputed: Google gives none, Amazon’s Mitra-v2 report says 1.6B.
• The top three aren’t statistically separated. Elo confidence intervals are wide at the top (Kumo-Tabular +150/-103, LimiX-2 +124/-88).

LimiX-2 sits second on the board, but with a warning. TabArena issue #648 (2026-10-07) showed it scoring a 1-AUC error of 0.0066 on the diabetes dataset, where every other model scored 0.15 or worse. When a tester reversed the column order, its error rose to 0.1536, in line with everyone else (tabarena#648). A model that truly learned from synthetic tables shouldn’t care about column order. One that memorised the real test labels would. TabArena added a warning on 2026-10-27; the LimiX team had not replied in the thread when the course’s research captured it.

The general lesson: a score far better than everyone else’s that collapses when you shuffle something irrelevant is the fingerprint of contamination. “Synthetic only” is a claim, not a guarantee.

TabArena is a “living” benchmark of 51 curated datasets (38 classification, 13 regression) from about 700 to 150,000 rows, with tuned and ensembled baselines (TabArena paper; TabFM blog). As of the 2026-10-27 board, the top 12 are eight in-context models plus four systems built around them (one, AutoGluon 1.6, bundles TabPFN-3). The best tree is about 600 Elo below the leader.

The slices tell a surprise. On small datasets, Kumo-Tabular and LimiX-2 tie at 1937, TabPFN-3.5 has 1814, and the best tree 1393. On medium datasets, Kumo-Tabular has 2426, TabPFN-3.5 2300, and tuned CatBoost 1613. The foundation-model lead is larger on medium data, which contradicts the 2025 view that these models only suit small tables (the course’s research read this from TabArena’s slice files).

Forecasting is in-context learning too: the history is the context. A forecaster reads your past values plus any covariates (extra series that help, such as temperature or a holiday calendar known ahead) and outputs future quantiles. Most models cut the history into patches, blocks of time steps embedded as one token each. Older models decoded autoregressively, one patch at a time, feeding each output back in; the newest emit the whole horizon in a single pass.

The message that started this course called TimesFM-3 “a 330M-parameter model that forecasts related time series at once with no task-specific fine-tuning and outputs the whole forecast in one forward pass”. Google released it on 2026-08-31 (Google Research). Each claim, checked:

It takes past-only and known-ahead covariates. Its downloadable weights are non-commercial; commercial use goes through BigQuery (google-research/timesfm). And it is less novel than the launch implied: Datadog’s Toto 2.0 used the same ingredients three months earlier (Toto 2.0 card).

Sources: Chronos-2 and card; Moirai 2.0; Toto 2.0 card; TTM-R3 card; Timestral; TimeH2O card. Latencies are the vendors’ own.

IBM’s TTM is the tiny-model case: its first version had about 1M parameters and pretrained in 24-30 hours on 6 A100s (TTM). TTM-R3 reaches a GIFT-Eval CRPS ratio of 0.520, better than the leakage-flagged Chronos-Bolt-Base (0.574) with about 150x fewer parameters. Timestral’s licence comes with a caveat: one report says it excludes firms above $20M monthly revenue (Let’s Data Science, 2026-10-19), a single source.

GIFT-Eval (Salesforce) covers 23 datasets, 144K series and 97 dataset-horizon settings across 7 domains and 10 frequencies. The course’s research recomputed these ratios from the raw result files (last updated 2026-10-23) and reproduced the published figures for TimesFM-3, Timestral and TimeH2O (GIFT-Eval Space).

Of 158 entries, 39 are “agentic” systems (routers or ensembles over several models), and they hold most top spots. Many were tuned or selected on the benchmark’s training split, so read the top of the board as “best system”, not “best model”.

fev-bench (Amazon) has 100 tasks across 7 domains, 46 of them with covariates, and bootstrapped confidence intervals (arXiv 2509.26468). Recomputed from the result files in autogluon/fev, fetched 2026-10-28:

How fast it moves: in May 2026, Prior Labs’ TabPFN-3 report said its TabPFN-TS-3 “ranks 2nd on fev-bench” (arXiv 2605.13986). Five months later it ranks about 20th.

Median seconds per 1,000 rows on TabArena (GPU for the foundation models, CPU for the trees):

A worked example. You serve 50 customers, each with a 10,000-row table. Assume the medians scale linearly (a simplification):

fit, tuned CatBoost:  1,346 s x 10 = 13,460 s per table (3.7 h)
                      x 50 tables  = about 187 hours
fit, TabPFN-3.5:      1.84 s x 10  = 18.4 s per table
                      x 50 tables  = about 15 minutes
predict 1M rows/day:  TabPFN-3.5   0.473 x 1,000 = 473 s
                      CatBoost default 0.025 x 1,000 = 25 s

The fit saving is about 700x. The prediction bill goes the other way: TabPFN-3.5 predicts about 19x slower than default CatBoost, on a GPU. Many small tasks with modest traffic favour the foundation model; one big table scored millions of times a day favours a tree, a cache or a distilled model.

In time series, the decoding style is the lever: Amazon’s Chronos-Bolt, which emits all quantiles at once, is “up to 250x faster” than the original token-by-token Chronos (card). General LLMs lose on cost: Tan and colleagues removed the LLM from three LLM-based forecasters, and accuracy matched or improved in 26 of 26 cases for Time-LLM. Time-LLM has about 6.6B parameters against 0.2M for its ablation, trained 3,003 minutes against 1.9 on one dataset, and inferred 28.2x slower (Tan et al., 2024-06-22). LLMs keep an edge only when the forecast hinges on free text about events or constraints (Context is Key).

Licences have become the business model. The leaders release weights for research and charge for production through hosted APIs (Prior Labs’ “Thinking” mode), cloud SQL functions (BigQuery’s AI.FORECAST and AI.PREDICT) and enterprise licences.

Sources: TabPFN-3, TabPFN-3.5, TabPFN README (weights need a login and licence acceptance), the model cards above. Prior Labs’ code is Apache-2.0; only the weights are restricted. Read the data rights lesson before you ship on any of these.

Both of these statements are true, and you need both:

• On broad benchmarks the gap is wide. The best single time-series models reach a CRPS ratio of about 0.44-0.46 against Auto-ARIMA’s 0.91 on GIFT-Eval, and fev-bench skill of about 0.50 against about 0.21 for a statistical ensemble.
• On long, clean, single-frequency business series, classical methods stay competitive. Nixtla’s ensemble of AutoARIMA, AutoETS, AutoCES and DynamicOptimizedTheta beat Chronos-T5-Large (710M) by 10% on CRPS and MASE across more than 50,000 M1, M3, M4 and Tourism series, about 5x faster on CPU (Nixtla). That is a competitor’s benchmark, undated. TimesFM’s own paper found seasonal ARIMA within statistical significance on the Darts datasets (TimesFM).

Where the baselines win in practice:

1. Large tables. TabArena stops near 150K rows. The Nature paper itself recommends CatBoost, XGBoost or AutoGluon for larger data and highly non-smooth regression; TabPFN-3.5’s authors say tuned MLPs lead on large grouped and temporal data.
2. High-volume prediction. Default CatBoost predicts about 19x faster, on a CPU.
3. Operations that need low error. Zero-shot TimesFM had 9.25% day-ahead solar error on Texas grid data; fine-tuned, 3.71% (arXiv 2604.22077). The vertical models lesson has the rest.
4. Stable, simple series. One study (read by the course’s research from a search summary, not the paper itself) found a simple baseline often beat every foundation model on electricity load in Singapore’s stable climate.

Where the foundation models clearly win: many varied tasks with no per-task ML effort, cold start and short histories (ForecastPFN targets series of about 40 or fewer points), covariate-rich forecasting (Chronos-2 reports its widest margins there), and small and medium tables, where a forward pass beats hours of tuning.

These are the APIs from each project’s README, checked by the course’s research on 2026-10-28. TabPFN returns a probability per class, the “typed decision plus probability” shape from Part 2.

# pip install tabpfn  (first run asks you to log in and accept the
# weights licence; TabPFN-3.5 weights are non-commercial)
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
from tabpfn import TabPFNClassifier

X, y = load_breast_cancer(return_X_y=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.3, random_state=0)

clf = TabPFNClassifier()           # default model: TabPFN-3.5 (tabpfn 9.x)
clf.fit(X_tr, y_tr)                # no gradient steps: the rows become context
proba = clf.predict_proba(X_te)    # forward pass over context + queries
print("ROC AUC:", roc_auc_score(y_te, proba[:, 1]))

# pip install chronos-forecasting 'pandas[pyarrow]'  (Apache-2.0)
# Shows two related meters forecast jointly, 24 hours ahead.
import numpy as np, pandas as pd
from chronos import Chronos2Pipeline

pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="cpu")

t = pd.date_range("2026-01-01", periods=24 * 28, freq="h")
context_df = pd.concat([
    pd.DataFrame({"id": sid, "timestamp": t,
                  "target": 10 + np.sin(2 * np.pi * np.arange(len(t)) / 24 + k)})
    for k, sid in enumerate(["meter_a", "meter_b"])
])
pred_df = pipeline.predict_df(
    context_df, prediction_length=24, quantile_levels=[0.1, 0.5, 0.9],
    id_column="id", timestamp_column="timestamp", target="target",
)
print(pred_df.head())

# pip install "timesfm[torch]"  (weights: non-commercial licence)
# Shows two series of different lengths forecast 12 steps ahead.
import numpy as np
from timesfm3 import TimesFM3Evaluator, ModelConfig

config = ModelConfig(checkpoint_path="google/timesfm-3.0-pytorch",
                     per_core_batch_size=32, device="cuda")
forecaster = TimesFM3Evaluator(config)

ts1 = np.linspace(0, 1, 100).astype(np.float32)
ts2 = np.sin(np.linspace(0, 24, 72)).astype(np.float32)
outputs = list(forecaster.predict_batch([ts1, ts2], horizon=12,
                                        return_quantiles=True,
                                        use_symmetric_averaging=False))
print(outputs[0].forecast.shape, outputs[0].quantiles.shape)  # (12,), (12, 9)

As of October 2026 the default designs are per-column embedding, row compression, then in-context learning over rows for tables (TabPFN-3, TabICL, TabFM), and a patched decoder alternating time and variate attention with single-pass quantiles for time series (TimesFM-3, Toto 2.0, Timestral). The open questions:

• Does real data matter? Kumo-Tabular, Mitra-v2 and TabICLv2 top the board synthetic-only; TabDPT says real data trains faster; Chronos-2 beats its synthetic-only twin by about 2%.
• Beyond 150K rows, every superiority claim is on vendor-internal data.
• Systems versus models. Ensembles tuned on benchmark training splits top GIFT-Eval. Nobody knows how well that transfers to production.
• Self-reporting. Almost every state-of-the-art claim here comes from the model’s maker; the independent anchors are the benchmarks’ raw files. Older models (Chronos, Chronos-Bolt, TimesFM 1.0 and 2.0, TTM-R1 and R2, Lag-Llama) also carry GIFT-Eval leakage flags.

• TabArena leaderboard: the live tabular board, with slices, timing and leak warnings.
• GIFT-Eval Space and fev-bench paper: the two forecasting benchmarks and their leakage rules.
• TimesFM-3 announcement: the architecture in Google’s words; compare it with the standings above.
• TabICLv2 (arXiv 2602.11139): the open, BSD-licensed model with released pretraining code.
• Nixtla’s Chronos experiment: the sharpest case for statistical baselines.
• Tan et al., are language models useful for forecasting? (arXiv 2406.16964): why a general LLM is the wrong tool for numbers.