March 1, 2026
Speed vs. Hype: place your bets
Show HN: Timber – Ollama for classical ML models, 336x faster than Python
Old-school AI gets a turbo button, and the Ollama comparison splits the crowd
TLDR: Timber promises microsecond-speed, compiled “classic” machine learning with an easy server setup. The crowd loves the speed but argues the Ollama comparison: you can’t swap fraud models like chatbots, and classical ML still matters even if generative AI steals the headlines.
Timber just rolled into Show HN promising a speed-demon for “classical” machine learning — the old-school, decision-tree stuff that powers fraud checks and credit scoring — and the crowd lit up. Fans cheered “I have been waiting for this!” while others loved the one-command, server-ready vibe, calling the concept “great.” The headline claim? 336× faster than Python thanks to compiled native code and microsecond latency. Translation: it’s meant to be blazingly quick and predictable, especially for teams fighting fraud on the front lines.
But the real fireworks ignited over the marketing: calling it “Ollama for classical ML.” One skeptic snapped, “Ollama is quite a bad example here,” arguing Ollama’s magic is easy model swapping, not compilation wizardry. Another practical voice pointed out that open-source fraud models aren’t plug-and-play anyway — you can’t just swap in a “generic” fraud brain and expect it to know your data. Meanwhile, a quieter chorus celebrated the reminder that not everything is chatbots: traditional ML still runs the world while generative AI hogs the spotlight. Cue memes about “finally something for the accountants” and “microsecond joy for people who live in spreadsheets.”
If you want receipts, Timber ships docs, an Ollama-style local API, and reproducible benchmarks — all right here: Docs. The vibe: speed thrills, marketing chills, and classical ML gets its moment back.
Key Points
- •Timber compiles tree-based ML models into optimized native C (C99) and serves them via a local HTTP API with microsecond-level latency.
- •It supports XGBoost (JSON), LightGBM (text), scikit-learn (pickle), ONNX TreeEnsemble, and CatBoost (JSON) model formats.
- •Benchmarks report up to 336× faster single-sample inference versus Python XGBoost, measured in-process on an Apple M2 Pro.
- •Comparisons highlight Timber’s zero runtime dependencies and small artifacts (~48 KB) versus alternatives like Python serving, ONNX Runtime, Treelite, and lleaves.
- •Limitations include narrow ONNX operator support (TreeEnsemble), CatBoost JSON-only exports, sklearn wrapper fragility, and the need to load only trusted pickles; roadmap targets embedded profiles and regulatory tooling.