An Indian face recognition model enters the global frontier

There are 497 milliseconds between an image entering Vogic and a face being matched.

In isolation, that number means very little. Until you ask what happens during those 497 milliseconds.

A machine receives pixels. It has to locate a face, represent its distinguishing features mathematically, compare that representation against another identity, and make a decision.

And it has to do this accurately enough that the decision can be trusted.

In July 2026, we submitted vogic-000 for evaluation by the U.S. National Institute of Standards and Technology.

It was our first submission.

The result puts an Indian-built model on the global accuracy-speed frontier.

99.57% verification accuracy. 497 milliseconds per face. One CPU core. No GPU.

Among the Indian models evaluated, none is simultaneously faster and more accurate. Globally, Vogic ranks 27th among the 132 evaluated models that run as fast as it or faster, according to our analysis of the NIST results.

The Indian models in the benchmark. No Indian model is both faster and more accurate than vogic-000.

Accuracy isn't enough

A face recognition system can be extraordinarily accurate and still be unusable in the physical world.

Imagine a city with thousands of cameras. An airport processing continuous passenger movement. A factory watching multiple access points. Or a security system operating at the edge with limited compute.

Every additional millisecond compounds.

Every additional GPU adds infrastructure, power, cost and dependency.

The real engineering problem therefore is not just about accuracy.

It is: How much intelligence can we produce, with how little compute, in how little time, without sacrificing the confidence of the decision?

This is why we care about the Pareto frontier.

A system sits on an accuracy-speed frontier when improving one dimension increasingly requires sacrificing the other. Vogic's first NIST submission puts it close to that frontier: only 26 models in the comparison outperform it simultaneously on both dimensions.

The global field. The shaded box holds the 26 models that beat vogic-000 on both speed and accuracy.

And there is still distance to travel.

The world's most accurate model in the comparison reaches approximately 99.80%, but takes roughly three times as long.

We know where the frontier is. Our job now is to move toward it.

What does 99.57% actually mean?

"99.57% accurate" sounds impressive but it is also an incomplete story.

NIST evaluates algorithms at defined false match rates. For this comparison, the operating point is one false match in 100,000 comparisons.

At that threshold, Vogic correctly matches 99.57% of genuine comparisons in the benchmark represented in our analysis.

That distinction matters.

In systems responsible for identity, access or public safety, an algorithm does not merely need to recognise the right person.

It must also know when not to recognise the wrong one.

NIST's evaluation methodology explicitly examines this trade-off between false matches and false non-matches across datasets and operating thresholds.

Why speed changes the system

497 milliseconds was measured on one Xeon CPU core, without a GPU. The benchmark also recorded zero failures to enrol.

This is not merely an optimization statistic.

It changes what can be built.

Models that require expensive centralized compute create one architecture. Models capable of operating efficiently closer to the source create another.

Cameras become more useful.

Edge deployments become more viable.

Infrastructure costs fall.

Response loops become shorter.

And intelligence can move closer to where events actually happen.

This is the deeper engineering philosophy behind Vogic.

Why should India care?

India operates at a scale few countries experience.

Its airports, railways, factories, borders, cities, defence installations and public infrastructure generate enormous quantities of physical-world data.

Increasingly, the systems interpreting that world will be intelligent, and a country cannot outsource the layers of its perception.

Sensors, models, compute, data, everything matters.

And the ability to integrate them into dependable systems matters.

This is what technological sovereignty looks like at the engineering level.

Not slogans about self-reliance, but Capability.

The first submission

There is another number from this benchmark that matters to us.

Or rather, 000.

This was vogic-000: our first NIST submission.

Several Indian models represented on the frontier reached their current positions after multiple submissions. Vogic entered the benchmark here.

So we don't see this benchmark as a victory lap.

We see a coordinate.

It tells us where we are.

The gap between 99.57% and the world's best tells us where to go next.

The 497 milliseconds tells us what we should refuse to sacrifice while getting there.

And an independent global benchmark gives our engineers something more valuable than applause:

a measurable frontier to attack.

vogic-001 is already aimed at it.

A note on speed

Updated 1 October 2026.

NIST says on its FRTE 1:1 results page that its recent speed figures are provisional:

The duration values for algorithms submitted since 2025-09-04 are provisional. From that date NIST ran all algorithms on a more recent Linux operating system (OS). The durations may be revised after the conclusion of our investigation into why numeric software is slower on identical hardware equipped with the new OS.

vogic-000 was submitted in July 2026, so its 497 milliseconds were measured on the new setup.

We expect about a 2x gain in speed in the next evaluation.

At about 250 milliseconds, with the same accuracy, the current version would sit very close to the top 5. Among the models at least that fast, only 3 to 6 would be more accurate, depending on how NIST revises the other recent runs.

It would also be close to the fastest models in the world at this accuracy. Of the 211 other models that are at least as accurate, only 3 to 6 would be faster. The fastest of them takes 190 milliseconds today.

Today, among the 42 models submitted in 2026 with at least our accuracy, vogic-000 is the 4th fastest. That puts it in the top 10%.

Protect what matters

At Pinaca, we believe that the physical world is becoming intelligent.

Cameras will cease to be passive recording devices. Sensors will cease to merely collect signals. Machines will increasingly perceive, reason and help humans make decisions.

Who builds and controls that intelligence will matter.

And which nations possess the capability to build it themselves will matter.

That is why a face recognition benchmark is, for us, about much more than face recognition. It is one small test of a much larger proposition:

Can world-frontier intelligence systems be built from India?

497 milliseconds.

99.57%.

First submission.

Our answer is beginning to take shape.

Protect what matters.