Compiled inference — not interpreted

ML models at
native speed.

Upload a tree-based model. We compile it to C99 and serve it in an isolated container. 50µs inference. No Python on the hot path.

Works with your stack

XGBoost
LightGBM
scikit-learn
CatBoost
ONNX
XGBoost
LightGBM
scikit-learn
CatBoost
ONNX

~50 µs

Median latency

1.38M

Rows/sec throughput

<2 ms

P99 under load

0 ms

Cold start

terminal

# upload → compile → serve in seconds

$ timber upload fraud-detector.json

✓ Parsed 200 trees · 50 features

✓ Compiled to C99 (6 optimisation passes)

✓ SHA-256: 8f3a…c2d1

$ timber serve fraud-detector

● Listening on http://localhost:11434

inference: ~120 µs/row · 625K rows/sec

Built for speed. Built for trust.

Everything you need to deploy tree-based models in production, without the Python overhead.

Compiled, not interpreted

Models are compiled to native C99 at deploy time. No Python runtime, no GIL, no garbage collector in the hot path.

Sub-millisecond inference

Single-row predictions in ~50 microseconds. Batch 10K rows in under 8ms. Consistent, predictable latency.

Isolated containers

Each deployment runs in its own container with dedicated resources. No noisy neighbors. Auto-scaling built in.

Audit-ready

Every compilation is SHA-256 hashed. Full audit trail. MISRA C-compliant generated code for regulated industries.

Supported frameworks

XGBoost
LightGBM
scikit-learn
CatBoost
ONNX

How we compare

Single-row inference, 50-tree XGBoost model.

RuntimeLatencyvs Timber
Python XGBoost~800 µs16×
Python sklearn~1,200 µs24×
ONNX Runtime~400 µs
TF Serving~2,500 µs50×
Timber~50 µs

Ready to deploy at native speed?

Upload your first model. See the difference in seconds.

Get started — it's free