TinyML Models

Larun uses a federation of specialized TinyML models, each optimized for a specific astronomical detection or analysis task. All models are under 100KB, enabling edge deployment and browser-based inference.

NEW: Cloud Platform

Try These Models Online

Upload FITS files and run instant inference with any of these 12 models in the cloud.

5 free analyses per month • No setup required • Results in <100ms

Download Models for Local Use

Get the complete model bundle for local inference on your own hardware (free forever).

Two Ways to Use These Models

Cloud Platform: Upload FITS files for instant inference (5 free/month, then $9/month for 50 analyses).View plans

Local Download: Download models and run unlimited inference on your own hardware (free forever).Get started

12
Trained Models
98.0%
Real Data Accuracy
<35KB
Avg Model Size
<20ms
Avg. Inference

Detection Models

Orbis

Trained

EXOPLANET-001

Transit detection, planet by planet

Primary exoplanet transit detection trained on real Kepler/TESS light curves. Feature-based classifier distinguishing transits, eclipsing binaries, and noise.

Accuracy: 98.0%Size: 43KB

Lumina

Trained

VSTAR-001

Every star has a rhythm

Variable star classification. Distinguishes Cepheids, RR Lyrae, Delta Scuti, and other variable star types.

Accuracy: 99.8%Size: 27KB

Helios

Trained

FLARE-001

Catch the star's outburst

Stellar flare detection. Feature-based classifier for quiescent, flare, and strong flare states.

Accuracy: 96.7%Size: 5KB

Gravitas

Trained

MICROLENS-001

Gravity bends light. We find it.

Microlensing event detection. Feature-based classifier for no event, simple lens, and complex events.

Accuracy: 99.4%Size: 5KB

Nova

Trained

SUPERNOVA-001

First to see the burst

Supernova and transient detection. Feature-based classifier for Type I, II supernovae, kilonovae, and TDEs.

Accuracy: 100.0%Size: 3KB

Discovery Federation Models

Try Citizen Discovery Engine →

Layer-2 server models running in the Citizen Discovery Engine. Inspired by Matteo Paz's VARnet which found 1.5M objects in NEOWISE — these models form the core of larun.space/discover.

Ripple

Trained

VARDET-001

Wavelet-powered variability detection

VARnet-inspired variability detector. Lomb-Scargle + Daubechies wavelet features fed into a Random Forest. Classifies NON_VARIABLE, TRANSIENT, PULSATOR, ECLIPSING — the same approach that found 1.5M objects in NEOWISE.

Accuracy: 97.2%Size: ~50KB

Sentinel

Unsupervised

ANOMALY-001

Nothing unusual escapes it

Isolation Forest anomaly detector on 14-dim feature vectors. Flags objects with unusual variability patterns — designed to catch Boyajian's Star analogs and other unexplained light curves.

Accuracy: Size: ~20KB

Clarity

Trained

DEBLEND-001

Separate what the sky mixed together

Detects blended / contaminated TESS pixels using multi-frequency analysis and pixel-crowding metrics (CROWDSAP, FLFRCSAP). Flags CLEAN, MILD_BLEND, STRONG_BLEND, CONTAMINATED.

Accuracy: 94.1%Size: ~15KB

Tempo

Algorithmic

PERIODOGRAM-001

Find the beat in the noise

Consensus period finder using 4 methods: Lomb-Scargle, BLS, Phase Dispersion Minimization, and Autocorrelation. Returns best period, confidence, and type (transit / pulsation / rotation / irregular).

Accuracy: Size: ~5KB
Try in DiscoveryNo weights needed

Analysis Models

Spectra

Trained

SPECTYPE-001

A star's fingerprint, classified

Stellar spectral classification. Classifies O, B, A, F, G, K, M, L type stars from photometric features.

Accuracy: 95.0%Size: 5KB

Resonance

Trained

ASTERO-001

Hear what stars are made of

Asteroseismology analysis. Detects solar-like, red giant, delta Scuti, and gamma Dor oscillations.

Accuracy: 99.8%Size: 21KB

Forma

Trained

GALAXY-001

Shape tells the story

Galaxy morphology classification. Feature-based classifier for elliptical, spiral, barred spiral, irregular, and merger galaxies.

Accuracy: 99.9%Size: 4KB

Run Models Locally

Use our TinyML models on your own hardware with pure NumPy - no TensorFlow required:

# Clone the repository
git clone https://github.com/Paddy1981/larun.git
cd larun && pip install numpy

# Load and run a model
from src.model import get_model

# Basic detection
detector = get_model("EXOPLANET-001")
detector.load("models/trained/EXOPLANET-001_weights.npz")
predictions, confidence = detector.predict(lightcurve_data)

# Multi-view detection (advanced)
from src.model.multiview_exoplanet import MultiViewExoplanetDetector
mv_detector = MultiViewExoplanetDetector()
mv_detector.load("models/trained/MULTIVIEW-EXOPLANET_weights.npz")
probs = mv_detector.forward(global_view, local_view, secondary_view)

Model Architecture

All Larun models share a common architecture optimized for TinyML deployment:

1

Input

Light curve data

2

Preprocessing

Detrend & normalize

3

Feature Extraction

1D Conv layers

4

Classification

Dense + dropout

5

Output

Probabilities