IBM has launched a “directed execution model” in IBM Quantum Compute Service, built on a new tool called Executor. IBM’s 5 October 2026 blog says it “runs your circuits exactly as directed” [1]. IBM’s documentation says its components are “currently in beta and might not be stable” [2].
Confirmed: IBM launched it (IBM blog and docs), as a beta. The performance figures below are company-reported; TSN found no independent benchmark.
What the terms mean
- Primitive. In Qiskit, IBM’s quantum software, a primitive is a standard type of job. Sampler returns a circuit’s measurement results. Estimator returns an average value of something you want to measure [1].
- Error mitigation. Quantum chips make mistakes. Mitigation runs many slightly varied copies of a circuit and processes the results to reduce the effect of noise. It does not fix errors inside the machine; error correction aims to.
- Directed. Until now, IBM’s servers chose how to vary and mitigate your circuits, a “black box” in IBM’s words [3]. Now you describe the variations on your own computer and Executor follows them. The docs say it “does not make any implicit decisions for you” [2].
What changes for developers
IBM released client-side Sampler and Estimator, rebuilt on Executor, in qiskit-ibm-runtime 0.50.0, which PyPI shows was uploaded on 24 September 2026 [4][5]. Their mitigation code now runs on your side, so you can inspect or change it [1].
For now they have their own import paths; IBM says the normal imports will switch over in the near future [3]. Jobs now report primitive_id as “executor” [3].
IBM’s blog says that in most cases only import statements change [1]. Its migration guide lists more: Estimator no longer learns noise implicitly for two methods, PEA and PEC, so you must run a separate step first; mixed shot values in one job are gone; and runs “might take longer” because more work happens on the client [3].
What IBM reports
In a blog example, IBM says “shaded lightcones” cut the sampling overhead (extra runs needed) of probabilistic error cancellation (PEC) “by roughly 3.4x while maintaining comparable accuracy” [1]. It also cites a study that ran distance-5 surface codes, a type of error-correcting code, on IBM’s Nighthawk chip and improved the logical error rate per round “by up to 2.8x” [1]. Both are company-reported.
What this does not show
- Stability. IBM’s docs say it may not be stable [2].
- Independent results. The 3.4x figure is one IBM example. TSN did not read the 2.8x study. No outside test is cited.
- Speed. IBM promises the “performance you have come to expect” [1] without numbers; its own guide warns of slower runs [3].
Sources
- IBM Quantum blog, “Directed execution: same performance, more control”, Jessie Yu, Tushar Mittal and Robert Davis, 5 October 2026 (company announcement). https://www.ibm.com/quantum/blog/directed-execution
- IBM Quantum documentation, “Directed execution model (beta)” (company documentation). https://quantum.cloud.ibm.com/docs/en/guides/directed-execution-model
- IBM Quantum documentation, “Migrate from server-side to client-side Sampler and Estimator” (company documentation). https://quantum.cloud.ibm.com/docs/en/guides/migrate-to-client-side-primitives
- PyPI, qiskit-ibm-runtime 0.50.0, uploaded 24 September 2026. https://pypi.org/project/qiskit-ibm-runtime/0.50.0/
- Qiskit/qiskit-ibm-runtime, GitHub release 0.50.0 changelog, published 24 September 2026. https://github.com/Qiskit/qiskit-ibm-runtime/releases/tag/0.50.0
Related stories
- Rutgers and IBM Report Nearly 5,000 Mid-Run Checks and Resets on a Chain of Up to 100 Qubits
- Quantum Hardware Check: DARPA’s Final Test Stage, Universal Quantum’s $100m Round and New Error-Correction Results (IBM in DARPA’s final stage)
- IonQ Says It Linked a Trapped-Ion Qubit to a Solid-State Memory at More Than 1,000 Entangled Pairs a Second
- NVIDIA ISING: The Open-Source AI Bridge to Practical Quantum Computing
