The Piestingtal
postl.ai is an independent AI laboratory in the upper Piestingtal. Its focus is open AI models that run entirely locally: the lab operates them on its own infrastructure, measures their capabilities, and proves them in daily use: 100% Linux, open source, without the cloud.
The lab
The lab operates open models locally and investigates their usefulness for code, text, media and automation.
Ryzen 9 9950X · RTX 5090 + RTX 3090 · 256 GB workstation RAM · Xeon inference server · ZFS storage
The local inference systems are owned and operated here in the Piestingtal. Model weights and the computing resources needed for these runs stay on-site.
Linux, open inference engines and accessible model weights make the setup inspectable. Model licenses differ; open weights do not automatically mean unrestricted use.
The local model endpoints process inference on our own hardware. Inference runs without an external API.
Models are evaluated in production use rather than synthetic demos: web infrastructure, autonomous agents, image and video production, automation, among others for supersauna.club, whose entire infrastructure is built and operated on these systems.
Fields of work
Six areas of work, with practical examples and illustrative animations.
Open language models from 8B to the trillion-parameter class. Quantisation choices, context setups, and balancing speed against quality for each task.
A sauna log that writes itself: record temperature and humidity, recognise an aufguss from the sensor curve and save the session. This illustrative coding example shows how a language model can turn a practical club task into a small local tool.
Outline an article for supersauna.club about the autumn season: training, the sauna tent and regional partners.
From topic to article outline: an opening, key sections and a closing for the existing autumn season article on supersauna.club. This illustrative planning example is based on the published text; GLM-5.3 is prepared as an experimental model.
The ComfyUI pipeline produces image, video with native audio, and music, from prompt to finished clip, entirely on in-house GPUs.
Source photoOpen video-diffusion model with a mixture-of-experts backbone. In image-to-video mode it animates a single still into a temporally coherent clip, inferring motion, expression and camera path frame by frame.
Latent diffusion transformer built for near real-time synthesis, with a natively generated audio track. It trades the multi-minute render of larger models for clips in seconds, at stable motion.
Original
FLUX.2Rectified-flow transformer for photoreal synthesis. In image-to-image mode it re-renders a reference under new light, style and atmosphere while holding composition and identity fixed.
Text-to-image model distilled to a handful of sampling steps and released under Apache-2.0, so genuinely free to run and modify. Notably strong at clean typography, emblems and logos.
Latent-diffusion model that generates a full song, vocals and backing, end to end in seconds. An open, self-hosted alternative to Suno. Press play for the first verse of the club song.
Object detection and tracking in live video, re-identification, plate and text recognition and segmentation, as building blocks for custom applications.
A single-stage detector: one forward pass finds and classifies every object in the frame. ByteTrack then links the detections across frames, so each object keeps its own ID, through occlusions and across the whole scene. That is the difference between counting boxes and counting things.
A box says roughly where. A mask says exactly which pixels. The same footage, segmented pixel by pixel: the basis for measuring, counting and cutting objects out of a scene rather than merely locating them.
From raw recordings and unlabelled measurements to something you can act on: transcription, clustering, semantic search across own archives. With sensitive data, local operation is not a preference but a prerequisite.
A recorded club meeting becomes text, and the text becomes fields: dates, decisions, open items, extracted into a form other systems can process. The audio never leaves the machine.
178 wines, 13 chemical measurements each, no labels at all. Reduced to two dimensions and handed to k-means, the samples sort themselves into three groups, and those groups turn out to be the three cultivars they came from. Nobody told the algorithm that; the structure was in the data. Real values, computed here.
The archive as a vector space: every document becomes a point, a question lands among them and pulls in the passages that actually answer it, with sources instead of guesses.
Training runs of its own, from LoRA adaptations for language models to classifiers and detectors on custom data. That includes the groundwork: cleaning, standardising, spotting outliers, splitting the set. A model is never better than the data it saw.
The same 178 wines, now with labels: a small neural network learns to tell the three cultivars apart from the 13 measurements. Two curves decide whether a run was any good: the loss has to fall, the accuracy on data the model never saw has to rise. Both curves below are the actual run, computed here.
Before a single epoch runs, the data has to be made comparable: measurements on wildly different scales (alcohol in percent, proline in milligrams) are standardised so no feature drowns out the rest. Outliers get flagged, and a quarter of the set is locked away so the model can be tested on data it has never met.
Autonomous agents with tool access take over recurring work (researching, monitoring, drafting, reporting) around the clock on in-house hardware.
Fritz lives in a Telegram chat. That is the whole interface. A sentence is enough: it looks up the sponsoring agreement, fills the club's invoice template and sends the finished PDF straight back into the chat. The model behind it runs on its own machine in the valley; no cloud service sees the request, the agreement or the invoice.
The fleet
Four systems for inference, infrastructure, model development and agent workflows. Inventory checked in September 2026; orakel retains its last documented July inventory.
Eight Xeon sockets and a terabyte of RAM for large language models. Hardware and model inventory last verified in July 2026; powered on as needed.
The infrastructure node for websites, storage and self-hosted services, including Nextcloud, Matrix and the lab’s web interfaces.
A Ryzen 9 workstation with 256 GB RAM, RTX 5090 and RTX 3090. The September Qwen benchmarks use the RTX 5090 with CPU/RAM offload; the 3090 is retained in the equipment inventory for its planned return.
The Super Sauna Club’s agent machine for automation, research and authorised security testing, with a Telegram interface.
Why local
Artificial intelligence is becoming basic infrastructure, as fundamental as electricity, water or the internet. And infrastructure is something you should be able to own: understand how it works, decide how it behaves, and be certain it will still be there tomorrow. Two principles therefore carry this lab:
Open models belong to no one exclusively. Their weights can be inspected, their behaviour can be studied and modified, and once obtained they remain available forever, regardless of what becomes of their maker.
What runs in-house keeps running, even when a service out there disappears, a price rises or a policy changes. The data never leaves the premises, the compute is predictable, and every result stays reproducible.
Open models, local inference and documented measurements — on our own hardware, in the valley.
A lab, not a hosting provider. Self-hosted compute infrastructure in the upper Piestingtal.
For questions and exchange · t.postl@pm.me