Skip to content
FICUS by Numascale, home

01 / 06 · 10 mAll dayToday

Flat by lunchtime: back in the case, and soon in a drawer.

On the charger

Round 01/07

Reflexes for Things · software now · silicon planned for 2028

Give things reflexes at nanojoules.

Situation awareness on the device's own chip. The main chip wakes only with a reason.

See the measurements
FICUS · Reflexes for Things · 1:06
Transcript · scenes are AI-generated illustrations; every number shown is a measured site figure

You notice. You understand. You anticipate. Before you even think… it's reflex.

Things have no reflexes. So they stay awake to decide, ship every reading away, and raise alarms nobody trusts. And every day, we recharge them.

FICUS gives things reflexes. A trained model, compiled into Reflex software: a white-box graph, on the device's own chip. Four point four nanojoules a decision. Measured.

Glasses that follow your gaze, and wake the main chip only with a reason. A hearing aid that notices the room change, and switches by itself. A motor that feels its bearing wear, and says why. A car that knows a relayed key is not you.

From your nightstand to orbit. Reflexes for Things. Every battery-powered device… its own situation awareness.

Discuss a device pilot →
Nearest charger

Scenes are AI-generated illustrations · evidence shown per scene

FIG. 01Measured Reflex comparison

Three measured studies

Aware on every sample, measured.

Trained models, compiled into Reflex software that decides on the device's own chip. AI a device has, not AI it calls.

● Toothbrush · same CPU
  1. FICUS4.4 nJ · 0.5 µs · 100%
  2. Neuton641 nJ · 80 µs · 100%
  3. TensorFlow Lite923 nJ · 109 µs · 24%
  4. 0 µsCPU time at the same ~8 mW110 µs

Done in 0.5 µs. The others keep the same CPU busy 79–108 µs longer.

Drawn to scale from measured averages · not a captured waveform1

Three separate studies · each with its own boundary2

FIG. 02No reflexes yet

Most things have no reflexes.

Your hand leaves a hot stove before your brain knows why. Smart glasses keep their main chip awake instead.

2–5h

smart-glasses battery, in a 16-hour day3

10mA

display-off draw, half of it the main chip

89%

sure decisions, FICUS glasses test

● Smart glasses · same frame, two ways to decide
Smart glasses in line art: the main processor in the temple glows hot orange and the battery drains.
Today: a large chip stays awake to decide whether anything is worth deciding.

Situation awareness of Things

  1. 01 · Notice

    One Reflex per signal, every sample.

    worn · speech · bearing fault · pupil X/Y

    Measured · nRF54LM20

  2. 02 · Understand

    Reflexes fused into one Instinct.

    worn + speech + head turn → in conversation

    Built · not yet on a product chip

  3. 03 · Anticipate

    How far a reading is from flipping the decision.

    more band-3 vibration → “inspect”

    A margin, not a forecast

  4. 04 · Wake with a reason

    Sure: act. Unsure: wake the main chip.

    89% sure in the glasses test

    Historical · glasses test

FIG. 03Energy per decision

7 workloads · Nordic nRF54LM20

Same CPU. Less energy.

One Cortex-M33, seven workloads. FICUS finishes first, so the chip sleeps sooner.4

36.7×

less energy than Neuton

84.1×

less energy than TensorFlow Lite

40.2×

faster than Neuton

86.0×

faster than TensorFlow Lite

Energy per call · each workload on its own linear scale · lower is better

BearingIndustrial condition monitoring

41.5× vs Neuton

FICUS
11.0 nJ
Neuton
457 nJ
TF Lite
850 nJ

ToothbrushWearables and personal care

145.3× vs Neuton

FICUS
4.4 nJ
Neuton
641 nJ
TF Lite
923 nJ

Magic WandWearables and personal care

65.6× vs Neuton

FICUS
12.9 nJ
Neuton
844 nJ
TF Lite
653 nJ

GearboxIndustrial condition monitoring

2.6× vs Neuton

FICUS
54.4 nJ
Neuton
142 nJ
TF Lite
590 nJ

TransportAsset tracking and logistics

20.3× vs Neuton

FICUS
28.8 nJ
Neuton
586 nJ
TF Lite
1,275 nJ

Eye trackingper pupil estimate

29.1× vs Neuton

FICUS
402 nJ
Neuton
11,720 nJ
TF Lite
54,688 nJ

MNIST digitsper digit · accuracy-matched

148.3× vs Neuton

lite12
41.4 nJ
Neuton
6,132 nJ
P3 fast
350 nJ
TF Lite
194,490 nJ

Measured · geometric mean of 7 · accuracy not averaged

FIG. 04Eye tracking

Eye tracking · Nov 2026

402 nJ per pupil estimate.

Pupil position from an event camera. Both FICUS profiles beat Neuton, Axon and TensorFlow Lite on time, energy and mean error.5

41–161×

faster than all three, energy-first

18.42px

mean error, accuracy-first

Energy per pupil estimatelinear · lower is better
FICUS energy-first0.40 µJ
FICUS accuracy-first0.92 µJ
Neuton11.7 µJ
Axon NPU12.7 µJ
TensorFlow Lite54.7 µJ
Time per pupil estimatelinear · lower is better
FICUS energy-first39 µs
FICUS accuracy-first90 µs
Neuton1,627 µs
Axon NPU1,602 µs
TensorFlow Lite6,352 µs
FIG. 05MNIST on the chip

MNIST digits · Nordic nRF54LM20B · Nov 2026

A digit in 4.8 µs.

Every rival meets the FICUS build at its own accuracy. FICUS is faster, leaner and more accurate, every time.6

41nJ

per digit · FICUS lite12, 94.76% accurate

Slow motion · MNIST · Nordic nRF54LM20B · 128 MHz1 ms on the chip = 0.25 s on screen

One image in. One digit out.Measured time per inference, to scale.

22.62 mschip time elapsed

Meanwhile

back-to-back digits

157digits from FICUS lite12while Neuton reads one
635digits from FICUS P3 fastwhile TensorFlow Lite reads one

One MNIST image on Nordic's nRF54LM20B: TensorFlow Lite CNN 22.62 ms, 98.94% accurate; Tsetlin Machine 9.10 ms, 97.38% accurate; Axon CNN 1.56 ms, 98.94% accurate; Neuton 759.6 µs, 92.82% accurate; FICUS P3 fast 35.6 µs, 99.18% accurate; FICUS lite12 4.8 µs, 94.76% accurate. FICUS lite12 reads 157 digits while Neuton reads one; FICUS P3 fast reads 635 digits while TensorFlow Lite reads one.

  • 148×

    less energy than Neuton

    FICUS lite12 · 157× faster · +1.94 points

  • 43.9×

    less energy than the Axon NPU

    FICUS P3 fast · 43.7× faster · +0.24 points

  • 555×

    less energy than TensorFlow Lite

    FICUS P3 fast · 636× faster · +0.24 points

  • 204×

    less energy than a Tsetlin Machine

    FICUS P3 fast · 256× faster · +1.80 points

FIG. 06What nanojoules buy

Things last

What a year of decisions costs.

Three identical coin cells, AI compute only. Move the decision rate.

AI-only budget · three identical CR2032 coin cells · 225 mAh × 3 V ≈ 2,430 JIllustrative · AI energy only

100/ s

One year in 6 seconds

One year365 days later

TF Lite ✕
JanAprJulOctDec

FICUS15.8 nJ / decision

1.58 µW average

2.1% of the budget in a year

Neuton460 nJ / decision

46.0 µW average

59.7% of the budget in a year

TensorFlow Lite826 nJ / decision

82.6 µW average

Empty on day 340

After one year at 100 decisions per second: FICUS has 97.9% left; Neuton has 40.3% left; TensorFlow Lite is empty on day 340.

Estimate · AI compute only · not device battery life7

Same energy: 100 decisions a second with FICUS. 3.4 with Neuton. 1.9 with TensorFlow Lite.

FIG. 07What a reflex unlocks

What a reflex unlocks.

16 worlds, from the bedside to orbit. Drag the line to compare.

On you, today: Five cables on one nightstand, every night, and the one you forget is flat by morning.
On you, with FICUS: The phone still charges. The brush, the ring and the buds skip the cable.

01 / 16 · On you

Things last · 1 m to the nearest charger

Your devices stop asking to be charged every night.

Reflexes → one Instinct · in-ear detect · wrist raise · sleep onset · brushing zone

TodayFive cables on one nightstand, every night, and the one you forget is flat by morning.

With FICUSThe phone still charges. The brush, the ring and the buds skip the cable.

To the nearest charger1 m

None of these objects calls a data centre to make its decision.

AI-generated scenes · evidence status shown per world8

Next step · Reflexes for Things

Give your product its first reflex.

Bring a model and a recording. We measure Reflex software on your chip.

Device makers

Pilot a reflex on your product

Two or three workloads, measured against your current model.

Start a pilot →

Engineers

Try Reflex Lab

Explore measured Reflexes. Profile your own CSV.

Open the Lab →

Investors

Investor materials

The round, the plan and the data room.

Request access →

Notes

  1. 1Toothbrush Reflex on the nRF54LM20 CPU, replayed 20,000× slower · illustration from measured averages, not a captured waveform · power × duration, drawn to scale · PPK2 · features in RAM → class · assumed constant 3.0 V · historical accuracy on this Reflex: FICUS 100%, Neuton 100%, TensorFlow Lite 24% · toothbrush is the widest Neuton margin of the five Reflexes; the seven-workload averages are in FIG. 03
  2. 2Three measured studies, each with its own boundary, never blended · the 7-workload average covers time and energy on the Nordic chip, never accuracy · details and caveats on Evidence
  3. 3Glasses battery: maker claims and reviews · 10 mA display-off draw: Halo, Brilliant Labs; about half attributed to the main chip · 89% sure decisions: FICUS glasses test, historical · each awareness step carries its own status: only Notice is measured on a product chip today
  4. 4Five Reflexes, eye tracking and MNIST digits against Neuton and TensorFlow Lite · nRF54LM20B DK, Cortex-M33 at 128 MHz, CPU only · all draw 7–10 mW while running · processed features in RAM → class (C1) · PPK2 charge × assumed constant 3.0 V · 784 windows · Nordic evidence v6.3, 4 Nov 2026 · eye tracking: prepared event-count map → X/Y, energy-first, operator-declared 3.0 V, Nov 2026 · MNIST: one image → digit, accuracy-matched (lite12 vs Neuton, P3 fast vs TensorFlow Lite), Nov 2026 · equal-weight geometric means of the 7 ratios; accuracy is not averaged (five-Reflex historical accuracy: FICUS 99.24%, Neuton 98.76%, TensorFlow Lite 85.58%)
  5. 5Separate run, Nov 2026 · nRF54LM20B, Cortex-M33 at 128 MHz; the Axon model runs on its NPU · prepared event-count map → X/Y, including each model's frontend · energy at the operator-declared 3.0 V · accuracy is descriptive (reused INI-30 cohort) · accuracy-first has better mean, median and P95 than all three · energy-first trades a little tail error: P95 61.8 px against 58.5 px for Axon · FICUS draws more power while running, 10.2 mW against 7.2–8.6 mW, and finishes far sooner · the sixth of seven workloads in the Nordic average, with energy-first
  6. 6nRF54LM20B DK, Cortex-M33 at 128 MHz; the Axon CNN runs on Nordic's NPU · one image → digit, separate firmware captures with identical input images, memory placements differ · energy at the 3.0 V setting · accuracy on the common 5,000-image validation set · each reference against the FICUS build at its accuracy level: lite12 (94.76%) against Neuton (92.82%), computed from the table; P3 fast (99.18%) against Axon, TensorFlow Lite and a CAIR Tsetlin Machine, factors as published · FICUS buys its speed with larger tables: lite12 holds 36 KiB of model payload against 16 KiB for Neuton and draws 8.57 mW against 8.07 mW; P3 fast holds 1,304 KiB; the saving comes from finishing sooner · payload excludes code, runtime and working RAM · the seventh workload in the Nordic average · an earlier host-CPU panel against ResNet-6 and ViT-4 is on Evidence
  7. 7Illustrative · equal-weight geometric means of the five measured Reflexes (15.8, 460 and 826 nJ per decision), compute only · processed features in RAM → class · PPK2 charge × assumed constant 3.0 V · excludes sensing, feature extraction, radio, idle current, losses and self-discharge · device battery life needs a pilot measurement · once deciding is nearly free, a Reflex's most valuable job is deciding when not to use the radio · Measurement method
  8. 8Scenes are AI-generated illustrations, not product photos · each world shows its evidence: measured, historical, pilot, exploring (nothing trained yet) or vision (not built) · sensing, radio and displays still draw power
  9. 9Measured: energy per decision on nRF54LM20, C1 boundary · product families are proposed pilot workflows, not customer deployments · device drawings are illustrations