Tessera Model Foundry

Edge models that fit, measured on the part.

A catalogue of twenty-one models, each with its flash footprint and inference time measured on the same physical board at 150 MHz, plus custom models we build, quantise and fit to your part.

Model	Task	Flash · bytes	ms
Vision · 11
mobilenet-imagenet-large	1000-class ImageNet recognition	1,535,296	1595.0
mobilenet-imagenet	1000-class ImageNet recognition	602,640	459.1
cifar10-resnet-large	10-class image classification	512,024	3474.6
vww-mlperf	Person present / absent	333,288	207.8
person-detect-vww	Person present / absent	300,568	198.3
ascii95	Printed character recognition	215,208	514.6
ascii95-ink	Printed characters, labelled by ink coverage	215,208	512.5
ascii95-seg	Printed characters, trained on segmenter crops	215,208	512.2
cifar10-resnet	10-class image classification	98,496	372.4
digits11	Digit recognition with a reject class	37,288	76.8
mnist-lstm	Handwritten digit recognition	13,952	5.9
Audio · 6
dtln-denoise	Speech noise suppression	372,720	30.9
toycar-anomaly	Machine-sound anomaly detection	276,976	33.1
streaming-wakeword	Continuous wake-word detection	74,520	22.9
kws-ds-cnn	12-way keyword spotting	53,936	61.9
micro-speech	Spoken yes / no	18,800	14.5
audio-frontend	Raw audio to spectrogram features	8,772	1.6
Language · 2
book-lm-large Proprietary	Text generation	5,388,144	1594.0
book-lm Proprietary	Text generation	642,288	195.3
Signal · 2
sine-regression	Regression, x to sin(x)	2,704	0.1
simple-add	Runtime floor measurement	976	9.7

Tessera supports open-source models: nineteen in this catalogue are under permissive open licences. The language models, the scene painter and the speech synthesiser are Tessera's own.

Generative

Scenes and speech, on a microcontroller.

Scene painter

A prompt becomes a 160×120 scene. The learned painter is 1,604 bytes, with 57,600 bytes of fixed buffers and no heap.

Generated scene: a house in front of mountains with clouds — a house in front of mountains with clouds

Generated scene: a train in a city at night — a train in a city at night

Generated scene: a boat on the ocean at night under stars — a boat on the ocean at night under stars

Generated scene: an elephant in a field near trees — an elephant in a field near trees

Speech

Text to speech by cascade formant synthesis. About 14 KB of code, no model file and under one percent of a part.

Bearing outer race fault detected. Confidence high.

Supply rail nominal. 12.4 volts.

System ready.

Find out where your devices can get smarter.

Book an opportunity assessment