Fluxnova benchmark results on one Amazon EKS environment.
One invoice-approval process, deployed unchanged on Camunda 7 and on Fluxnova, and also run on Camunda 8, on Amazon EKS in June 2026. Zero failures on every engine at every volume. The interesting result is not which one won.
Preliminary results showing relative behavior on the documented workload. This is not a certified production benchmark, and it is our process on our infrastructure; your models are not our models. Full method and environment specifications are below and available in full on request.
Two engines in one class. One in another.
Fluxnova and Camunda 7 completed a thousand workflows within a few seconds of each other. Camunda 8 took fifteen to nineteen times as long on the same process, which is consistent with a distributed architecture that adds per-task overhead.
Fluxnova Con, Wednesday 4 November. It runs inside Open Source in Finance Forum New York, day 1, at Convene Brookfield Place. Ryan Johnston of Summit58 and Rob Stevens of iTmethods present “Fluxnova: Enterprise Performance & Camunda 7 Compatibility” on the event schedule. See the week →
Every number, at both volumes.
Throughput is completed instances divided by total wall-clock time. Both volumes ran on the same driver, the same invoice process and the same out-of-the-box configuration.
| Engine | 300 workflows | 1,000 workflows | Failures |
|---|---|---|---|
| Camunda 7 v7.24.0 | 6.7s · 45/s | 16.7s · 60/s | None |
| Fluxnova v2.0.x | 6.0s · 50/s | 21.3s · 47/s | None |
| Camunda 8 v8.8.0 | 104s · 3/s | 315s · 3/s | None |
Every run completed every instance. The difference between these engines is fit for this workload, not reliability.
What we read from the results.
Fluxnova was ahead at 300 workflows and 4.6 seconds behind at 1,000. If you are deciding whether to move off Camunda 7 Community Edition, read that as the same performance class on this process, not a gain or a loss.
On one process at two volumes, we do not read that as a meaningful advantage for either. Anyone presenting one of those two results on its own is selecting a result.
Some organizations need exactly that. The question this benchmark cannot answer for you is whether those properties are worth the overhead on your own processes.
No engine here is unreliable. If a benchmark of this shape had produced failures, that would have been the story instead.
One benchmark on one process is evidence about that process. It is a reason to run your own, not a substitute for having done so.
An architectural difference, stated without a verdict.
The three engines are not three implementations of one design, and the numbers above are mostly a picture of that.
What this means for a Camunda 7 estate. On this process Fluxnova ran in the same performance class as Camunda 7, and the invoice model was deployed unchanged. Compatibility and operating effort for your own models are what the assessment and the thirty-day environment are for. Run it on your own volumes →
This ran on the 2.0 line. 3.0 changed what runs inside a process.
These figures were produced in June 2026 against Fluxnova 2.0.x, and they are a throughput measurement of a deterministic invoice process. Version 3.0 shipped in July with MCP integration and agentic adhoc subprocesses, which means an agent can now invoke a process and determine its own steps at runtime. That is a change in what the engine can do, not in how fast it does it, and this benchmark says nothing about it.
How it was run, so you can take it apart.
Published so you can judge how far it carries to your own models. A benchmark whose method is not stated is an assertion with a number attached.
| What | How |
|---|---|
| The process | One BPMN invoice-approval definition, deployed unchanged to Camunda 7 and to Fluxnova, and also run on Camunda 8. No tuning of the model between Camunda 7 and Fluxnova. |
| The environment | Amazon EKS, us-east-1. Out-of-the-box configuration on every engine, so none of them is tuned and none is handicapped. |
| The driver | 300 and 1,000 instances started across ten concurrent threads, measuring wall-clock time until every instance completes. Throughput is completed instances divided by that total. |
| The versions | Camunda 7 v7.24.0, Fluxnova 2.0.x, Camunda 8 v8.8.0. Run June 2026. Fluxnova 2.0 was released as v2.0.0 on 15 April and v2.0.1 on 14 May 2026 (FINOS release list). |
| Who ran it | iTmethods, in June 2026, on our own Amazon EKS environment. |
| What it is not | Not a certified production benchmark, not tuned for peak on any engine, and not a projection for your estate. Full environment specifications on request. |
Our numbers are a reason to get your own.
We provision a dedicated Managed Fluxnova environment for your team for thirty days, at no charge and with no commitment. Bring this invoice benchmark or bring your own BPMN models.
Your models, your volumes, your data, and the results are yours to keep whether or not the conversation goes anywhere.
Tell us what you want to test.
We will provision a dedicated Managed Fluxnova environment and come back to you with access details and the next steps.
Discuss a benchmark using your workload.
The edition you are on, roughly how many process definitions are in production, and what is forcing the move. We will come back with what the migration actually involves, and what we would want to measure first on your own models rather than ours.