qmesh

The quantum substrate of the Neura Parse stack.

One IR. Every modality. Provenance by default. Fault-tolerant ready.

Website · Docs · Roadmap · Architecture · Examples

---

Why qmesh, why 2026

Quantum computing crossed three thresholds in the last twelve months:

• Logical qubits outperform physical qubits. Quantinuum H2 reached Microsoft Level 2 Resilient with logical error rates 800× below physical. IBM's roadmap puts Kookaburra (4,158 qubits across three chips) into 2026. IonQ committed to 80,000 logical qubits via modular trapped-ion interconnects.
• Modality boundaries dissolved. Photonic interconnects between trapped-ion and neutral-atom systems shipped commercially in April 2026 (IonQ, Cisco UQS). Neutral-atom platforms are Spectrum's "2026 big leap."
• The toolchain didn't follow. Qiskit, Cirq, PennyLane and friends remain provider-centric. Cross-modality programs require bespoke glue. Auditability is an afterthought.

qmesh is the missing layer: a single intermediate representation that crosses gate, neutral-atom, photonic-CV and pulse modalities; an AI-augmented compilation path; a fault-tolerant promotion that's one flag away; and a runtime where every result is signed, chained, and audit-ready by default.

---

The Neura Parse stack

flowchart TB
    subgraph Apps["Applications"]
        QANTIS["<b>QANTIS</b><br/>Quantum navigation"]
        QMANN["<b>QMANN</b><br/>Quantum-augmented memory"]
        NowFlow["<b>NowFlow</b><br/>Agentic workflows"]
    end
    subgraph OS["Operating Layer"]
        NeuralOS["<b>NeuralOS</b><br/>AI-native embedded Linux"]
    end
    subgraph Q["Quantum Substrate"]
        qmesh["<b>qmesh</b><br/>Modality-agnostic IR · FT mode · Provenance · DAG scheduler · AI copilot"]
    end
    subgraph HW["Hardware"]
        SC["Superconducting<br/>(IBM Heron, Google Willow)"]
        TI["Trapped-ion<br/>(Quantinuum Helios, IonQ)"]
        NA["Neutral-atom<br/>(QuEra, Pasqal, Atom Computing)"]
        PH["Photonic-CV<br/>(Xanadu, PsiQuantum)"]
    end

    QANTIS --> qmesh
    QMANN --> qmesh
    NowFlow --> qmesh
    NeuralOS -.coordinates.-> qmesh
    qmesh --> SC
    qmesh --> TI
    qmesh --> NA
    qmesh --> PH

Loading

qmesh is the foundational layer that QANTIS (quantum navigation, IBM Heron-validated) and QMANN (quantum-memory neural networks) compile through. New Neura Parse products that touch a QPU enter the stack at this line.

---

Six pillars

01 — Modality-agnostic IR A single typed IR (qmesh.ir) for gate, neutral-atom, photonic-CV and pulse programs. Deterministic content-hash. Cross-frontend semantic equivalence — write once, run on any backend the program is compatible with.	02 — Fault-tolerant mode Surface, repetition, BB qLDPC and cultivation-aware lattice surgery — promotion is a single config flag. PyMatching, sliding-window, BP+OSD and neural decoders are interchangeable. Resource estimates in the Microsoft RE shape, signed end-to-end.	03 — Modality reach Pasqal Pulser, Strawberry Fields, MrMustard, Bloqade frontends. QutipEmulator, Aer, Stim, OpenPulse backends. First-class ChannelOp lets a single program cross modalities — gate → Rydberg → CV in one signed run.
04 — Hybrid DAG scheduler Typed DAG nodes (QPUPrimitive, ClassicalTask, MCMRegion, Barrier, Fanout, EntanglementBarrier). Topological executor with thread-pool parallelism. SLURM / PBS connectors. Crash-safe resume via signed parent_manifest_hash lineage.	05 — AI copilot Intent compiler for common patterns. LLM provider abstraction (Ollama / Anthropic / OpenAI / Mock). Token-level grammar masking with real-time GrammarMaskedProvider. Neural decoder training pipeline (Stim → PyTorch → signed weights). QuanBench evaluation suite.	06 — Provenance & compliance Every run produces an ed25519-signed manifest. Tamper-evident compliance packs with offline-verifiable hash chains. Cert-chain trust model with TrustStore. Metriq export for public benchmarking. Designed for regulated industries.

---

Quick start

pip install -e .[all]

# 1 — Hello, quantum
python3 examples/hello_bell_crossbackend.py

# 2 — A real algorithm
python3 examples/vqe_h2.py                    # E = −1.859 vs FCI −1.857

# 3 — Fault-tolerant memory
python3 examples/ft_logical_memory.py         # Λ = 23 across d=3 → d=5

# 4 — Cross-modality, single signed run
python3 examples/dag_cross_modality.py        # gate → Rydberg → CV

# 5 — Compliance pack for auditors
python3 examples/compliance_export.py         # signed chain, offline-verifiable

# Service mode
uvicorn qmesh.service.app:app --port 8788

# CLI (16 commands)
qmesh submit circuit.qasm --shots 4096
qmesh check ledger/2026-05-02/<hash>.json
qmesh compliance-pack ./ledger --since 2026-05-01

---

Built for regulated industries

qmesh is the quantum layer for the same five sectors Neura Parse builds for:

Sector	What qmesh enables
Defense	Cert-chain signed runs, tamper-evident audit packs, classical+quantum DAG with SLURM/PBS
Aerospace	Quantum navigation (QANTIS) compiled through qmesh; reproducible, manifest-signed flight pipelines
Industrial	Hybrid optimisation pipelines with HPC connectors; Mitiq ZNE mitigation for noisy-era runs
Healthcare	Compliance pack with offline-verifiable hash chain; cert-chain trust model satisfies regulatory provenance
Automotive	Sensor-fusion + quantum-accelerated planning via cross-modality DAGs

---

Composes — does not replace

qmesh is not a Qiskit or PennyLane killer. It's the layer above them.

Layer	qmesh uses
Frontends	Qiskit · Cirq · OpenQASM 3 · Pulser · Strawberry Fields · MrMustard · Bloqade · PennyLane · qBraid
Compilation	TKET · BQSKit · MQT
Mitigation	Mitiq (ZNE)
Simulators	Aer · Stim · QutipEmulator · Strawberry Fields
Decoders	PyMatching · BP+OSD (ldpc) · custom neural / transformer
LLMs	Ollama · Anthropic · OpenAI · Mock fallback
Reporting	Metriq

The same teams that ship those projects win when qmesh wins. That is the design intent.

---

Validated against the 2026 baseline

• Bell, GHZ (3-25 qubits), QFT, VQE on H₂ match analytical / FCI results to 4 decimals.
• Surface-code memory at distance 3 → 5 reproduces the published Λ ≈ 23 break-even.
• Cultivation parameters match Gidney & Shutty (arXiv:2409.17595): 4×10⁻¹¹ logical T-state error at 5×10⁻⁴ physical noise.
• Cross-modality programs run on three modalities × four backends in a single signed manifest.
• Test suite: 157 passing, 1 skipped (tiktoken-optional), in 9.6 seconds.
• All 20 examples in examples/ execute end-to-end on the 2026 stack below.

Verified 2026 dependency matrix

Versions qmesh is tested against (Python 3.10–3.13):

Package	Tested	PyPI latest	Notes
qiskit	2.4.1	2.4.x	Pulse module removed in 2.0; OpenPulse backend handles it
qiskit-aer	0.17.2	0.17.2	Released Feb 2026
cirq	1.5.0	1.6.x
stim	1.15.0	1.15.0	Required for FT mode
pymatching	2.3.1	2.3.x	v2.3 brought correlated matching
pulser	1.7.2	1.7.x	Pasqal Rydberg simulator
strawberryfields	0.23.0	0.23.0	Maintenance; MrMustard is the successor
mrmustard	0.7.3	0.7.x	Xanadu's differentiable CV simulator
bloqade	0.33.0	0.33.x	QuEra neutral-atom
pennylane	0.42.3	0.42.x
qbraid	0.12.0	0.12.x	Requires ≥0.11 for API V2
pytket	2.16.0	2.16.x	TKET compiler bindings
mitiq	0.47.0	0.47.0	ZNE error mitigation
ldpc	2.4.1	2.4.x	Optional, BP+OSD decoder

Full per-area research notes live in docs/research/.

---

Roadmap

timeline
    title qmesh delivery timeline
    Phase 1 — α : Modality-agnostic IR : Frontends · Backends · Router : Provenance · Service mode
    Phase 2 — α/β/γ/δ : Surface · Repetition · BB qLDPC : Lattice surgery · Cultivation circuit : Litinski merge-CNOT observable
    Phase 3 — α/β/γ/δ : Pulser · Strawberry Fields · MrMustard : ChannelOp first-class : OpenPulse backend skeleton
    Phase 4 — α/β : Hybrid DAG scheduler : SLURM / PBS connectors : Crash-safe resume + entanglement barrier
    Phase 5 — α/β/γ/δ : Intent compiler · LLM providers : Transformer decoder · Grammar masking : Real-time GrammarMaskedProvider · QuanBench
    Phase 6 — α/β : Metriq export · Tamper-evident pack : Cert-chain trust model · PennyLane / qBraid plugins

Loading

What's still gated on external research or vendor APIs:

• AlphaQubit-2-scale neural decoder (100M+ parameters, 10⁶ Stim shots — research-grade)
• Live Metriq HTTP submission (Unitary Foundation endpoint contract)
• IonQ photonic / Cisco UQS middleware (vendor APIs, not yet public)
• Calibrated DRAG + cross-resonance pulse synthesis (vendor calibration data)
• CRL / OCSP cert revocation (validity windows ship today)

The full plan with citations: PLAN.md.

---

Inside Neura Parse

qmesh is one component of the Neura Parse intelligence stack:

Layer	Project	Status
Quantum substrate	qmesh	This repo
Quantum applications	QANTIS, QMANN	Public
AI-native OS	NeuralOS	Public
Agentic workflows	NowFlow	Product
Project management	TaskNebula	MIT

Build the intelligence stack, ship autonomous products. — Neura Parse

---

Resources

• Architecture deep-dive — ARCHITECTURE.md
• Roadmap with citations — PLAN.md
• Research notes — docs/research/ (hardware, software, simulation, QEC, gaps, AI/hybrid)
• Examples — examples/ (20+ runnable demos)
• Tests — tests/ (157 passing, 9.6s)

---

Community

• Website — neuraparse.com
• GitHub — github.com/neuraparse
• Issues & contributions — github.com/neuraparse/qmesh/issues

---

License

Apache 2.0 — see LICENSE.

────────

_{Built with research-backed engineering at Neura Parse.
© 2026 Neura Parse. Three layers. One intelligence.}