AI OpenFree

Posted on Oct 6

VIDRAFT Selected to Use Quantinuum's Nexus and H-Series: What It Means When an AI Lab Gets Real Quantum Hardware

#ai #quantum #llm #research

TL;DR

VIDRAFT was selected, after a review by Quantinuum, to use the full-stack Nexus quantum computing platform and H-Series ion-trap quantum computers under an evaluation allocation, with hands-on support from Quantinuum engineers.
Quantinuum has held the world Quantum Volume record since 2021. In September 2025 its 56-qubit H2 system reached a Quantum Volume of 33,554,432 (2 to the power of 25).
The review did not come out of nowhere. VIDRAFT had already run symmetric-key cryptanalysis on a real IBM Heron processor and built an AI solver that tackles the 2D Hubbard model on a single GPU.
This post is the engineering context behind the selection: what we ran on real quantum hardware, why a classical AI solver still matters, and what ion traps change for the roadmap.

Why would a quantum hardware vendor hand an AI lab an evaluation allocation?

Access to top-tier quantum hardware is not something you simply buy. The best machines are allocation-limited even for paying customers, and vendors are selective about who gets evaluation time because onboarding costs them engineering hours.

So the interesting question is not "VIDRAFT got access." It is "what did VIDRAFT show that justified the access." The answer is a track record on real quantum hardware plus a classical AI method that addresses a problem quantum machines are supposed to own. Those two lines of work are what this article unpacks.

What did VIDRAFT actually run on a real quantum computer?

Our quantum research team ran a key-recovery experiment against the Even-Mansour construction, a minimal symmetric-key cipher structure, on a real quantum computer built on IBM's latest Heron processor (the ibm_kingston backend).

The headline engineering result: previously published runs on real hardware topped out around 4 bits (N=4). We pushed the recovered key size to 10 bits (N=10), and validated five distinct attacks across four cipher variants.

Two details matter for anyone who works with near-term quantum devices:

Real hardware, not a simulator. Noise, readout error, and limited coherence time are the whole challenge. A clean simulator result at N=10 is routine. A hardware result is not, because circuit depth and qubit count fight against decoherence.
It was picked up by editors, not by a press release. The paper was selected for SemiEngineering's weekly security research review, alongside work from Meta, Google, Radboud University (Netherlands), and Politecnico di Milano (Italy). An editorial board chose it. We did not pitch it.

That is the credibility signal. When you ask a vendor for scarce hardware time, "we already shipped a real-hardware result that an independent editorial board flagged" is a stronger argument than any slide.

If you have quantum hardware, why build a classical AI solver?

This is the part engineers usually push back on, so let us be direct. Quantum hardware today is noisy and small. For many strongly correlated physics problems you cannot yet get a trustworthy answer from a quantum device alone. So VIDRAFT built a classical solver that uses AI to close the gap.

The solver is Δ-Engine (Delta-Engine), a strongly correlated quantum solver. Its target is the 2D Hubbard model, the standard minimal model for high-temperature superconductivity and a benchmark that has resisted exact solution for decades.

The engineering claim is specific and falsifiable: Δ-Engine validates the doped 2D Hubbard model on a single GPU, with no supercomputer and no quantum computer in the loop. The method has two layers:

A neural quantum state (NQS) produces a precise reference value for the ground-state energy.
A cheaper approximate calculation runs fast, and the AI corrects its error against the NQS reference.

We certified results with a confidence grade up to a 6x6 lattice. The grading matters: an uncertified energy number for a correlated system is close to useless, because you cannot tell a real result from a variational artifact.

A short, illustrative sketch of the two-layer idea (pseudocode, not production):

# illustrative only
def delta_engine(lattice, hopping_t, coulomb_u):
    e_ref = neural_quantum_state(lattice, hopping_t, coulomb_u)  # precise, slower
    e_cheap = approximate_solver(lattice, hopping_t, coulomb_u)  # fast, biased
    correction = ai_error_model(e_cheap, lattice, hopping_t, coulomb_u)
    e_final = e_cheap + correction
    grade = confidence_grade(e_final, e_ref)   # certify or reject
    return e_final, grade

So the two research lines are complementary, not redundant. Use the real quantum device where it already wins. Use the AI solver where the device cannot yet be trusted. The selection to use H-Series is about pushing the first line further while the second line keeps the physics honest.

What is Quantinuum's H-Series, and what does an ion trap change?

Quantinuum's H-Series machines are trapped-ion quantum computers. Ion traps differ from superconducting qubits (the family IBM's Heron belongs to) in ways that matter for the kind of deep, structured circuits cryptanalysis needs:

All-to-all connectivity. Any qubit can interact with any other, so you avoid the SWAP overhead that eats circuit depth on fixed-layout superconducting chips.
High gate fidelity and long coherence. That is largely why Quantinuum has led Quantum Volume, a single aggregate metric of usable circuit size, since 2021.
The 2-to-the-25 record. In September 2025 the 56-qubit H2 reached Quantum Volume 33,554,432, a world record (source: Quantum Computing Report).

For our Even-Mansour work, connectivity and depth are the binding constraints. Moving from a superconducting backend to a high-fidelity ion trap is exactly the kind of platform change that could let the N=10 result grow.

How does this fit VIDRAFT's broader stack?

VIDRAFT organizes its work around three axes: AI as the brain, quantum as the microscopic world, and physical AI as the macroscopic world. The quantum research sits on the second axis.

On the first axis, our open 180B model Darwin-180B-RSI tops ten official Hugging Face leaderboards. The thesis is to combine that AI capability with quantum methods that reach down to the level of materials and molecules, so the same organization can attack drug discovery, new materials, and physics problems with one coherent stack.

Related quantum work already underway includes quantum error correction (QEC) research, a materials simulation platform (MaterialsOS), and diagnostics that evaluate how resistant a cipher is to quantum attack. The Quantinuum selection widens the hardware side of that program from IBM superconducting machines to the world's leading ion-trap systems.

FAQ

Did VIDRAFT break a real cipher on a quantum computer?

No, and that distinction matters. The Even-Mansour work is a research-scale key-recovery experiment that extends real-hardware results from N=4 to N=10. Ten-bit keys are not fielded cryptography. The value is methodological: showing that a deeper, structured cryptanalytic circuit runs on real hardware and validates across multiple attacks and variants.

Why not just use a quantum computer for the Hubbard model instead of Δ-Engine?

Because near-term quantum devices cannot yet give a trustworthy ground-state energy for a doped 2D Hubbard lattice at the sizes that matter. Δ-Engine delivers a certified answer today on a single GPU. The two approaches are kept side by side so the classical solver can check the physics while the hardware matures.

What is Quantum Volume, in one sentence?

Quantum Volume is a single number that captures the largest square circuit (equal width and depth) a machine can run reliably, so it folds qubit count, connectivity, and error rates into one comparable metric.

Is the SemiEngineering selection a paid placement?

No. It was an editorial selection by SemiEngineering's weekly security research review, which also featured work from Meta, Google, Radboud University, and Politecnico di Milano. VIDRAFT did not issue a press release for it; the editors chose it.

What happens after the evaluation period?

VIDRAFT and Quantinuum agreed to hold follow-up discussions on the scale of future projects and further collaboration once the evaluation allocation is used. Research results obtained on the Quantinuum platform will be disclosed in stages, following the agreed process with Quantinuum.

Which hardware families is VIDRAFT now working across?

Both main qubit families: IBM's superconducting Heron (where the Even-Mansour experiments ran) and Quantinuum's trapped-ion H-Series (the newly granted evaluation platform). Running the same research questions on two physically different architectures is itself a useful cross-check.

DEV Community