Zero-shot machine learning for tabular data, inside DuckDB
This extension runs tabular foundation models, TabPFN-style in-context learners, so classification and regression become a single SQL statement. No Python, no training loop, no MLOps: the model reads your labelled rows as context and predicts the rest.
SELECT customer_id, yhat AS will_churn, yhat_score
FROM tabfm_classify('history', 'churned', test := 'prospects', keep := ['customer_id'], model := 'mitra');Built for AI pair-programming. anofox-tabfm ships a Claude Code plugin: five skills with the full, checked SQL surface, so Claude writes working
tabfm_*queries from a plain-English request. Install with/plugin marketplace add DataZooDE/anofox-tabfm→/plugin install anofox-tabfm@anofox-tabfm.
Built with StableAI LimiX. The
limix-2mmodel is StableAI's LimiX-2M, used under the Stable AI Technology Co., Ltd. License v1.0 (licence). Several models are non-commercial: check thecommercialcolumn in Models before you build on one.
- Classification and regression in one call:
tabfm_classify,tabfm_regress - Two forms: an explicit train / test pair, or one table whose missing targets are the rows to score
- Probabilities, not just labels:
yhat_scoreand aprobamap per row
- Built in, selectable by name: Mitra and Mitra-v2, TabDPT v1.2 and v1.3, TabPFN v2 / 2.5 / 2.5-real / 2.6 / 3 / 3.5 / 3.5-Fast, TabICL v2, Causilo, LimiX-2M, Nori-30M, EXAONE Tabular, Kumo-Tabular (small, medium, large), Orion-BiX, Orion-MSP and Google TabFM v1 / 1.1
- Your own model, registered entirely in SQL with
tabfm_register_model - One SQL surface: swap
model :=to compare models on your own data
- Synthetic rows that keep the relationships between columns:
tabfm_generate - Fill NULL cells, conditioned on the rest of each row:
tabfm_impute
- Metrics as SQL aggregates: accuracy, F1, ROC-AUC, log loss, ECE, RMSE, MAE, R², and more
- Leakage-safe cross-validation:
tabfm_cross_validate - Plots through anofox-visualization: model comparison, confusion matrix, ROC and calibration curves
- CPU on Linux, macOS (arm64) and Windows
- GPU with one call,
CALL tabfm_accelerate(): NVIDIA CUDA, AMD ROCm, Apple MLX
- No weights shipped: you download each model's weights from Hugging Face yourself
- Licence gates: a model whose licence needs an acknowledgement will not download without one
- Errors that say what to run
- Installation
- Quick Start
- Models
- Claude Code skills
- Documentation
- Status
- Development
- License
- Telemetry
- Feedback
INSTALL anofox_tabfm FROM community;
LOAD anofox_tabfm;Latest builds: the community build can trail a release by days. The DataZoo repository has it the
moment it is out, but its builds are unsigned: start DuckDB with duckdb -unsigned, and install httpfs
first (why):
INSTALL httpfs; LOAD httpfs;
SET custom_extension_repository = 'https://get.anofox.com';
INSTALL anofox_tabfm;
LOAD anofox_tabfm;Check which build you got:
SELECT extension_version FROM duckdb_extensions() WHERE extension_name = 'anofox_tabfm';-- 1. Weights are fetched over HTTPS, once: mitra is Apache-2.0, ungated, about 303 MB
INSTALL httpfs; LOAD httpfs;
CALL tabfm_download('classification', model := 'mitra');
-- 2. iris, read from Hugging Face, with every fifth species hidden as NULL
CREATE TABLE iris AS
SELECT SepalLengthCm AS sepal_len, SepalWidthCm AS sepal_wid,
PetalLengthCm AS petal_len, PetalWidthCm AS petal_wid,
CASE WHEN row_number() OVER () % 5 = 0 THEN NULL ELSE Species END AS species
FROM 'hf://datasets/scikit-learn/iris/**/*.csv';
-- 3. Rows with a species are the context; predict the NULL ones, zero-shot
SELECT sepal_len, petal_len, yhat AS predicted_species, yhat_score
FROM tabfm_classify('iris', 'species', model := 'mitra')
WHERE species IS NULL;Every predict, generate / impute and weight-lifecycle function takes model :=; cross-validation uses the
default model. Downloading does not choose a model for you: with no model := and no
SET anofox_tabfm_default_model = 'mitra', a call fails and says how to pick one. Regression needs its own
weights: CALL tabfm_download('regression', model := 'mitra');. The full walk-through with output:
Getting Started; runnable files in examples/.
| Model | By | Tasks | Licence (tabfm_list_models()) |
Commercial | Gated |
|---|---|---|---|---|---|
mitra |
AWS AutoGluon | classify, regress | apache-2.0 |
yes | no |
mitra-v2 |
AWS AutoGluon | classify, regress | apache-2.0 |
yes | no |
tabdpt |
Layer 6 AI | classify, regress | apache-2.0 |
yes | no |
tabdpt-v1-3 |
Layer 6 AI | classify, regress | apache-2.0 |
yes | no |
tabicl-v2 |
Inria (soda-inria) | classify, regress | bsd-3-clause |
yes | no |
orion-bix |
Lexsi Labs | classify | mit |
yes | no |
orion-msp |
Lexsi Labs | classify | mit |
yes | no |
nori-30m |
Synthefy | regress | apache-2.0 |
yes | no |
kumo-tabular-small |
NVIDIA | classify, regress | openmdw-1.1 (keep the agreement and NVIDIA's notice when you distribute) |
yes | no |
kumo-tabular-medium |
NVIDIA | classify, regress | openmdw-1.1 |
yes | no |
kumo-tabular-large |
NVIDIA | classify, regress | openmdw-1.1 |
yes | no |
tabpfn-v2 |
Prior Labs | classify, regress | apache-2.0 (Prior Labs License v1.1: Apache-2.0 + attribution) |
yes | yes |
limix-2m |
StableAI | classify, regress | limix-2m-license-v1.0 (Apache-2.0 + attribution: "Built with StableAI LimiX") |
yes | yes |
tabpfn-v2-5 |
Prior Labs | classify, regress | tabpfn-2.5-license-v1.1 |
no | yes |
tabpfn-v2-5-real |
Prior Labs | classify, regress | tabpfn-2.5-license-v1.1 |
no | yes |
tabpfn-v2-6 |
Prior Labs | classify, regress | tabpfn-2.6-license-v1.0 |
no | yes |
tabpfn-v3 |
Prior Labs | classify, regress | tabpfn-3-license-v1.0 |
no | yes |
tabpfn-v3-5 |
Prior Labs | classify, regress | tabpfn-3-5-license-v1.0 (no hosted service) |
no | yes |
tabpfn-v3-5-fast |
Prior Labs | classify, regress | tabpfn-3-5-license-v1.0 (no hosted service) |
no | yes |
causilo |
Nums AI | classify, regress | causilo-license-v1.0 (hosted, API or SaaS use needs a separate licence) |
no | yes |
exaone-tabular |
LG AI Research | classify, regress | exaone-nc-1.2 |
no | yes |
tabfm-v1 |
classify, regress | tabfm-non-commercial-v1.0 |
no | yes | |
tabfm-v1-1 |
classify, regress | tabfm-non-commercial-v1.0 |
no | yes |
Gated means the download needs SET anofox_tabfm_accept_hf_license = true; first, your acknowledgement
of that licence. Some Hugging Face repositories also need account approval and a token, which is a separate
step: Getting Started.
Which model to pick, and how they compared: Choosing a model. Sizes,
backends and attribution texts: Models and weights.
Memory on the CPU. mitra's row attention grows with the square of the context: about 3 GB per 1,000
context rows at 15 features, so 2,000 rows peak near 7 GB and 4,000 exceed 15 GB, well below the default
anofox_tabfm_max_rows (10,000). Sample the context to fit your RAM (#107).
Tabular ML is easier when your AI assistant knows the API. anofox-tabfm ships an installable
Claude Code plugin: five skills that give Claude every function, named
parameter, opts key, output column, model and error message of this release. Ask "predict churn for the
prospects from history with a commercially usable model, then cross-validate it" and Claude writes the
tabfm_classify and tabfm_cross_validate calls.
/plugin marketplace add DataZooDE/anofox-tabfm
/plugin install anofox-tabfm@anofox-tabfm
| Skill | Covers |
|---|---|
anofox-tabfm-predict |
tabfm_classify / tabfm_regress: both forms, features, opts, output columns, segments |
anofox-tabfm-models |
the 23 models, licences, the licence gate, the weight lifecycle, your own model, the context cache |
anofox-tabfm-evaluate |
metrics, scoring rules, cross-validation, the _by plotting macros |
anofox-tabfm-generate |
synthetic rows and imputation |
anofox-tabfm-gpu |
CUDA, ROCm and MLX plugins, devices, precision, the ROCm precompile |
The skills live in plugins/anofox-tabfm/ and are checked against the built
extension on every pull request: each SQL example runs, and each function, setting and model name must
exist (how). In this repository, load them
directly with claude --plugin-dir ./plugins/anofox-tabfm.
| Guide | Description |
|---|---|
| Getting Started | Install, download a model, licence gates, a first prediction on iris |
| Choosing a model | Selecting a model, the default, your own model, how the models compared |
| Evaluate and compare | A churn walk-through, cross-validation, plotting the results |
| GPU acceleration | tabfm_accelerate(), checking which device served a query, the ROCm first compile |
| Generate and impute | Synthetic rows and filling NULL cells |
| Claude Code skills | Installing the plugin, what each skill covers, how they stay in sync |
| Topic | Description |
|---|---|
| Functions | tabfm_classify, tabfm_regress, missing targets, segments, opts, output columns |
| Models and weights | Every model, download / load / unload / remove, the cache, secrets, tabfm_register_model |
| Generation | tabfm_generate, tabfm_impute, options, outputs, cost |
| GPU acceleration | CUDA, ROCm and MLX plugins, the support matrix, precision, platform limits |
| Settings | Every setting with its default, and the opts keys |
| Troubleshooting | Error messages, what they mean, what to run |
| Evaluation | Metrics, cross-validation, scoring rules |
| Plotting | The six _by macros for anofox-visualization |
The full index, including design notes and the history of plans and reviews, is docs/README.md.
| Document | What it is |
|---|---|
| docs/REAL_MODELS.md | Reference: each real model, how it was ported and verified, licences, gated repositories |
| docs/GENERATE.md | Reference: how generation and imputation work, and what binning costs |
| docs/BUILDING.md | Reference: building and testing from source |
| docs/DYNAMIC_BACKENDS.md, docs/KV_CACHE_DESIGN.md | Design: GPU backends as runtime plugins; the context cache |
| docs/ROCM_SINGLE_EVAL_POS.md, docs/MLX_SPIKE_RESULTS.md | Report: how models that read the split from len(y) were given masked ROCm graphs; the MLX spike |
| docs/README.md | History: the plans and reviews behind each feature |
| CHANGELOG.md | Every release |
Moved sections. The old README's sections now live here (their links still land on this table):
| Old section | Now |
|---|---|
| Quickstart | Quick Start, Getting Started |
| Classification and regression | Functions |
| Generate synthetic data and fill gaps | Generate and impute, Generation |
| The registry, managing weights | Choosing a model, Models and weights |
| Worked example, evaluation, plotting | Evaluate and compare, Evaluation, Plotting |
| Settings | Settings |
| Flavors (CPU / GPU), Status & scope | GPU acceleration, GPU |
| Building | docs/BUILDING.md |
The extension is calendar-versioned (2026.10.26 is the twenty-sixth of October 2026) and is built and
tested on Linux (amd64, arm64), macOS (arm64) and Windows (amd64) against DuckDB 1.5.6, and on DuckDB 2.0.
Every model runs a single estimator (n_estimators > 1 is not available yet). What each GPU backend serves:
the support matrix.
git clone --recurse-submodules https://github.com/DataZooDE/anofox-tabfm.git
cd anofox-tabfm
GEN=ninja make release # needs vcpkg: see docs/BUILDING.md
make test_release # sqllogictests + C++ unit tests (fixture models, no real weights)
make check-docs # every documented SQL example and name, docs and Claude Code skillsPrerequisites, the GPU plugins and source-build flavors: docs/BUILDING.md.
- This extension's code: MIT, see LICENSE.
- Model weights: the extension ships only weight-free graphs and never redistributes weights. Each
model's weights are your own Hugging Face download, under that model's licence, listed in
Models and in
tabfm_list_models()(license,commercial, and the notice to keep inattribution). The repository and test fixtures contain zero model weight bytes (fixtures are random-init). - Non-commercial models:
tabfm-v1,tabfm-v1-1,tabpfn-v2-5,tabpfn-v2-5-real,tabpfn-v2-6,tabpfn-v3,tabpfn-v3-5,tabpfn-v3-5-fast,exaone-tabularandcausilo.causilo's licence also requires a separate licence for hosted, API or SaaS use. - Bundled graphs: the Kumo-Tabular graphs encode NVIDIA's architecture, which OpenMDW-1.1 counts as part of the Model Materials, so the binary carries that agreement and NVIDIA's notice; see THIRD_PARTY_NOTICES.md.
- Attribution:
tabpfn-v2is under the Prior Labs License v1.1 (Apache-2.0 plus attribution).limix-2mis under the Stable AI Technology Co., Ltd. License v1.0 (Apache-2.0 plus Section 10): if you make available a product or service that incorporates it, link the licence and display "Built with StableAI LimiX"; a model derived from its weights must start its name with "LimiX". Details: docs/REAL_MODELS.md.
Licences are as published by each project at the time of writing; this is not legal advice.
Sends anonymous usage telemetry (extension load and per-function call counts: no data, no queries, no
identifiers beyond a hashed machine id) to PostHog EU, matching the other anofox extensions. CI environments
are auto-detected and telemetry is disabled there. Opt out any time with export DATAZOO_DISABLE_TELEMETRY=1
or SET anofox_telemetry_enabled = false;. What is collected: TELEMETRY.md.
If anofox_tabfm misbehaves — a model that will not load, a prediction that looks wrong, a device it will
not use — please open an issue. Hardware, drivers and
hub availability differ in ways we cannot reproduce here, so a report with your setup is the fastest path to
a fix. Errors from the model functions end with that link.
If it saved you time, a star on the repo helps other people find it.
The first time you load the extension in an interactive terminal each day, a small banner says the same
thing. It never prints when output is piped, in notebooks, or in CI. Silence it with
SET datazoo_banner = false; or DATAZOO_NO_BANNER=1.