Show HN: Timber – Ollama for classical ML models, 336x faster than Python

Ollama for classical ML models.

Timber compiles trained tree-based models (XGBoost, LightGBM, scikit-learn, CatBoost, ONNX) into optimized native C and serves them over a local HTTP API.

• No Python runtime in the inference hot path
• Native latency (microseconds)
• One command to load, one command to serve

📚 Docs: https://kossisoroyce.github.io/timber/

Timber is built for teams that need fast, predictable, portable inference:

• Fraud/risk teams running classical models in low-latency transaction paths
• Edge/IoT teams deploying models to gateways and embedded devices
• Regulated industries (finance, healthcare, automotive) needing deterministic artifacts and audit trails
• Platform/infra teams replacing Python model-serving overhead with native binaries

pip install timber-compiler

# Load any supported model (auto-detected)
timber load model.json --name fraud-detector

# Serve it (Ollama-style workflow)
timber serve fraud-detector

curl http://localhost:11434/api/predict \
  -d '{"model": "fraud-detector", "inputs": [[1.0, 2.0, 3.0, ...]]}'

The 336× claim is measured against Python XGBoost single-sample inference.

• Hardware: Apple M2 Pro, 16 GB RAM, macOS (recorded by script)
• Model: XGBoost binary classifier, 50 trees, max depth 4, 30 features
• Dataset: breast_cancer (sklearn)
• Warmup: 1,000 iterations
• Timed: 10,000 single-sample predictions
• Metric: in-process latency (not HTTP/network round-trip)
• Baseline: Python XGBoost (booster.predict)

See benchmarks/ for:

• run_benchmarks.py (Timber vs Python XGBoost + optional ONNX Runtime/Treelite/lleaves)
• system_info.py (hardware/software metadata)
• render_table.py (markdown table output)

Run:

python benchmarks/run_benchmarks.py --output benchmarks/results.json
python benchmarks/render_table.py --input benchmarks/results.json

• ONNX support is currently focused on TreeEnsembleClassifier/Regressor operators.
• CatBoost support expects JSON exports (not native binary formats).
• scikit-learn parser supports major tree estimators and pipelines; uncommon/custom estimator wrappers may fail.
• Pickle parsing follows Python pickle semantics — only load trusted artifacts.
• XGBoost support is JSON-model based. Binary booster formats are not the primary input path.
• Optional benchmark backends (ONNX Runtime, Treelite, lleaves) are skipped unless installed/configured.

• Improve framework/version compatibility coverage (including more edge-case model exports)
• Broaden ONNX operator support beyond tree ensembles
• Strengthen embedded deployment profiles (ARM Cortex-M / RISC-V presets)
• Add richer benchmark matrices and public reproducibility reports
• Expand safety/regulatory tooling around audit + MISRA-C workflows

End-to-end runnable examples live in examples/:

• quickstart_xgboost.py
• quickstart_lightgbm.py
• quickstart_sklearn.py

They generate model files you can load immediately with timber load.

Timber includes a full technical paper: paper/timber_paper.pdf

@misc{royce2026timber,
  title        = {Timber: Compiling Classical Machine Learning Models to Native Inference Binaries},
  author       = {Kossiso Royce},
  year         = {2026},
  howpublished = {GitHub repository and technical paper},
  institution  = {Electricsheep Africa},
  url          = {https://github.com/kossisoroyce/timber}
}

• Contributing guide: CONTRIBUTING.md
• Code of conduct: CODE_OF_CONDUCT.md
• Security policy: SECURITY.md

pip install -e ".[dev]"
pytest tests/ -v

Apache-2.0