Timber, an ahead-of-time compiler for tree-based machine learning models, has launched with claims of 336x faster inference compared to Python XGBoost implementations. Created by Kossiso Royce and Electricsheep Africa, the project compiles trained models into self-contained C99 binaries with zero runtime dependencies.
Performance Metrics and Supported Frameworks
The compiler achieves approximately 2 microseconds single-sample latency with throughput reaching 500,000 predictions per second. Generated artifacts measure around 48 KB, making them suitable for resource-constrained environments where Python runtimes are unavailable or impractical.
Timber supports major ML frameworks including:
- XGBoost (JSON format, all objectives)
- LightGBM (text and binary formats)
- scikit-learn (pickle format for major estimators and pipelines)
- CatBoost (JSON export only)
- ONNX (TreeEnsembleClassifier and TreeEnsembleRegressor operators)
Five-Stage Compilation Pipeline
The compiler operates through a systematic five-stage process. First, it parses models into a framework-agnostic intermediate representation. Second, it applies optimizations including dead-leaf elimination, threshold quantization, and branch sorting. Third, it generates portable C99 code without dynamic allocation or recursion. Fourth, it compiles the code via gcc or clang to produce a shared library. Finally, it serves predictions through an Ollama-compatible HTTP API.
The generated C99 code avoids recursion and dynamic memory allocation entirely, ensuring predictable performance characteristics. This design makes Timber particularly suitable for edge inference and latency-critical applications. Related work on optimizing compilers for decision trees provides additional academic context.
Production Deployment Advantages
Timber positions itself as "Ollama for classical ML models," emphasizing simplicity with "one command to load, one command to serve." The project includes a 146-test suite and comprehensive documentation, alongside a technical paper explaining its implementation.
The repository, launched February 27, 2026, has accumulated 523 stars across 25 commits. Primary development uses Python, with generated output in C99 for maximum portability.
Addressing ML Deployment Pain Points
While deep learning frameworks dominate headlines, tree-based models remain widely deployed in production systems. Timber addresses the persistent challenge of slow Python inference for these classical ML approaches, particularly relevant for high-throughput serving scenarios.
The compiler's elimination of runtime dependencies simplifies deployment in containerized environments and edge devices. The 48 KB binary size contrasts sharply with Python environments requiring hundreds of megabytes.
Key Takeaways
- Timber compiles tree-based ML models to C99 with claimed 336x speedup over Python XGBoost
- Generated binaries achieve 2 microseconds latency and 500,000 predictions per second throughput
- Supports XGBoost, LightGBM, scikit-learn, CatBoost, and ONNX model formats
- Generated C99 code avoids recursion and dynamic memory allocation for predictable performance
- 48 KB binary size and zero runtime dependencies enable deployment in resource-constrained environments