The rzfz.ai Stack: local matters
Five ways local AI is changing how we build and test software
The renaissance of local intelligence
We live in a tension field: generative AI keeps getting bigger, faster and more capable — while the EU AI Act tightens the regulatory net around it. For companies, that means the question is no longer just what’s technically possible, but what’s legally permitted, ethically defensible and economically sound. With the rzfz.ai Stack, we build toward exactly that answer — the intelligence has to come to the data, not the other way around.
The technical foundation: unified memory
Classic workstations keep system RAM and video memory strictly separate, so models quickly hit the ceiling of the graphics card. The rzfz.ai Box instead runs on AMD Strix Halo with 128 GB of unified memory, of which around 96 GB is usable for models (extendable to 110 GB) — CPU and GPU share the same memory pool. That makes it possible to run models that simply don’t fit on a classic desktop GPU, at a power draw of 30 to 300 watts — quiet enough for a normal office, no dedicated server room required.
On top of that, the stack orchestrates inference through GPUStack with llama.cpp, keeps knowledge vectorized in PostgreSQL with pgvector, and gives every user a single interface through Open WebUI. Dify handles workflow automation — the glue that connects models to tools like Gitea or your own ticketing system.
Five use cases for development and testing
Reasoning-capable models like our default, qwen3.6, don’t just change chat — they change how we build and verify software:
- Requirements architect — A requirements document, as PDF or plain text, goes into Open WebUI; a Dify workflow hands it to the model with instructions to identify actors, use cases and gaps. The result is a diagram plus a list of open questions — before a single line of code exists.
- The relentless test designer — Instead of testing only the happy path, the model formally applies equivalence partitioning and boundary value analysis, generates edge-case test data, and produces a test matrix you can export into your existing test-management tool.
- Legacy code refactoring — OpenHands (currently an experimental module of the stack), our autonomous coding agent, reads old components straight out of Gitea, explains the business logic they contain, writes regression tests for the existing code first, and only then proposes a refactor. The knowledge locked in that legacy code never leaves the local network.
- Autonomous bug hunter — On every push, OpenHands (experimental, yet already in practical use at customer sites) analyzes the diff, looks for logical errors rather than mere syntax mistakes, and comments on suspicious spots directly on the merge request — a pair programmer that never sleeps.
- Living documentation — On a regular cadence, the agent compares code against existing documentation and submits updates as a pull request. Onboarding new team members gets noticeably faster as a result.
Thinking it through economically
The reflex reaction — “the cloud is cheaper, I only pay for what I use” — falls short, especially for reasoning models, which generate internal thinking tokens on every iteration that the cloud bills per request. The Box, by contrast, is a one-time investment: whether a model gets asked ten times a day or ten thousand, the running cost barely moves. What that means for your organization specifically is what the Configurator works out for you.
The takeaway: local matters
“Local matters” isn’t a slogan for us — it’s the program: compliance isn’t an innovation blocker, it’s the precondition for winning back freedoms that got quietly given up in the cloud. Fittingly, the joint project with our customer PSA on the rzfz.ai Stack won the Constantinus Award 2026 — 1st place, category Standardsoftware und Cloud Services.
First published in SEQIS QualityNews H2/2025.