Strip out the vocabulary and WiMi Hologram Cloud's announcement comes down to a single claim. A computer model told one handwritten number apart from another.
That is the task. MNIST is a collection of 70,000 scanned handwritten digits, each one a small grid of grey pixels. It has served as the entry-level test for image recognition since the late 1990s. Binary classification means the model deals with two digits at a time, such as separating every three from every eight.
Standard software handles this without difficulty. A basic classical neural network scores above 99% accuracy on the full 10-digit version, running on a laptop.
What the hybrid does
H-QNN splits the job in two.
The image is shrunk first. Hundreds of pixel values are compressed into a much shorter list of numbers. Those numbers are then loaded into a quantum circuit, written in as the amplitude and rotation angle of each qubit.
Inside the circuit, rotation gates and entanglement gates mix the values together. The result comes back out as ordinary numbers. A classical neural network layer then makes the final call.
So the quantum section does feature extraction. The classical section does the decision and the training.
That division of labour is the whole design. Today's quantum machines are small and noisy. Keeping the quantum stage shallow and handing everything else to classical code is the standard workaround across the field.
The word missing from the announcement is hardware
WiMi says the next milestone is moving from simulation towards experiments on real quantum hardware. That tells you what the current work is. A quantum circuit modelled on classical computers.
Simulating a small number of qubits on a normal machine is routine. It shows whether the maths holds. It shows nothing about whether the approach survives contact with physical qubits, where noise and decoherence wreck fragile quantum states.
One release in a long series
This is not a one-off. WiMi has put out quantum machine learning research announcements through 2026 covering quantum convolutional networks, generative models, federated training, visual tracking and data encoding schemes. Each follows a similar shape. A new architecture, a benchmark result, a statement about future potential.
The company's core business is holographic and augmented reality products, from in-vehicle head-up displays to LiDAR. The quantum work sits alongside that as research output rather than revenue.
What would count as progress
Three things.
Results on a dataset harder than MNIST, which the wider research community treats as a starting point rather than a proving ground.
Results on physical qubits rather than a simulator.
A measured advantage over a classical model of similar size. Not a claim of improved efficiency, a like-for-like comparison with the numbers shown.
Until one of those arrives, the plain reading holds. A research team built a hybrid model, ran it in software, and reported that it worked on the easiest test in the field.
