42

The model decides how much to think. Adaptive-depth inference that matches accuracy while cutting compute and energy 20–83%.

---

What it is

42 is an inference framework that allocates computation per input: easy inputs exit early, hard inputs run deeper. The halting policy is trained end-to-end with the network, so the savings are accuracy-preserving — not the usual accuracy-for-speed trade-off. 42 runs as software on existing NVIDIA GPUs; no custom silicon, no change to your data pipeline.

It comes in two forms: a neural variant (deep-learning / language workloads) and a tree variant (structured / tabular data).

Results (public benchmarks)

Benchmark	Samples	42	Best competitor	Savings
CERN Higgs (A100)	8.5M	0.837 ROC-AUC	XGBoost ~0.80	82.6% energy
Forest Covertype	581,012	87.75% acc	LightGBM 85.75%	71.8% compute
Credit-card fraud	284,807	0.757 PR-AUC	XGBoost 0.836	60.3% compute
NSL-KDD intrusion	148,517	73.67% acc	CNN-LSTM 69.75%	47.2% compute
MNIST	60,000	98.0–98.1%	XGBoost 97.4%	adaptive
Kaggle S6E2 (4,370 teams)	external	0.95497 AUC — top 1%	top 0.95535	server-side

Kaggle Playground S6E2 is server-side scored with no access to test labels, so test-set leakage is structurally impossible. Details in RESULTS.md.

Key properties

• Accuracy-preserving savings — the model decides for itself when it has done enough.
• Capacity preserved — the full stack still exists for the hard cases; this is not pruning, distillation, or quantization.
• Free drift signal — per-input compute utilization is observable at runtime, an out-of-distribution monitor for free.
• GPU-native — runs on standard NVIDIA hardware; deployable today.

Illustrative API

Interface shown for illustration. This repository contains no implementation, model equations, or parameters.

from fortytwo import Model          # illustrative

model = Model(task="classification")
model.fit(X_train, y_train)
preds  = model.predict(X_test)      # adaptive depth per input
report = model.energy_report()      # joules & layers used per prediction

Docs

overview · benchmarks & methodology · FAQ

How the numbers are kept honest

Every published result is produced under a versioned, anti-fabrication protocol: deterministic splits with automatic leakage and overfit gates, directly-measured GPU energy, and hash-stamped, timestamped receipts. 42 runs one fixed configuration on every dataset while each competitor gets a full hyperparameter search — a comparison deliberately generous to the baselines.

Citation

@misc{42-2026,
  title  = {42: Adaptive-Depth Inference},
  year   = {2026},
  note   = {Patent pending. https://github.com/42-global67/42}
}

Status & contact

Patent pending. The core mechanism is not published here. For evaluation, reproduction, or partnership, open an issue on this repository.

© 2026. All rights reserved.