LADX
Start free
← Resources
AI·8 min read·

Running a local model for PLC work

Most OT networks have no internet access, which makes local models the only option rather than a preference.

Short answer

Yes, and on an OT network it is often the only permitted option. A machine with 16 GB of RAM runs a capable 7B to 14B parameter model through Ollama or LM Studio, exposing an OpenAI-compatible endpoint that most tools can point at. Smaller models are noticeably weaker at code, which is why a validation loop around the output matters more locally than it does in the cloud.

The loop that makes it usableModel writesany providerCompiler answerspass / failYou read itonly if it passedfailures go back, with the errorsthis is what makes a free model good enough

The reason to care about local models in automation is not privacy in the abstract. It is that the network the work happens on frequently has no route to the internet, by design.

What runs on what

  • 16 GB RAM, no GPU. A 7B model runs, slowly. Usable for explanation and small generations.
  • 12 GB VRAM or more. A 14B model runs comfortably and is genuinely useful for code.
  • 24 GB or more. Larger models, or several people sharing one endpoint.

Ollama and LM Studio both expose an OpenAI-compatible API, which means anything that can point at a base URL can use them, including LADX.

The honest quality gap

A local 14B model is weaker than a frontier hosted model, especially on instructions it has seen less of. Pretending otherwise wastes people's time.

What closes most of the gap is not a bigger model but a loop: compile the output, feed the errors back, try again. Quality moves from the model to the checking, and the checking is deterministic.

The deployment question nobody asks early enough

Who patches it. A model server on an OT network is a service with a lifecycle, and 'the engineer who set it up has left' is the usual failure mode. Treat it like any other machine component: documented, imaged, and owned by somebody.

Common questions

What hardware do I need to run a local model for code work?
16 GB of RAM runs a 7B to 14B parameter model acceptably on CPU, slowly. A GPU with 12 GB or more of VRAM makes it comfortable. For a shared engineering workstation, a single mid-range GPU serves a small team through an OpenAI-compatible endpoint.
Are local models good enough for PLC code?
On their own, weaker than the large hosted models, particularly on less common instructions. With a compile-and-repair loop around them the gap narrows sharply, because most of the quality comes from the checking rather than from the first attempt.

Keep reading