The Humanizer plug-in is a small Stratat extension that rewrites generated text so it reads smoothly while preserving meaning, and it works entirely on your own computer using the Strata engine inside the Stratat client. It is intended for US users who need local privacy and fast iteration on large models without relying on remote inference.
Why the Humanizer exists
The core idea is simple: when you draft a document, email, or note in Stratat, the model sometimes outputs phrasing that feels stilted or repetitive. Humanizer automatically applies gentle smoothing that keeps your tone and facts intact while reducing awkward loops. Because the model runs fully on-device, no text ever leaves your machine, which matters for personal data and compliance.
How to set up the Humanizer add‑on
The plug‑in is distributed as a static module that loads together with the Strata engine when you open Stratat. After installing the latest Strata model build that includes the humanizing weights, simply enable the feature from Stratat's settings panel under Extensions > Humanizer. No extra drivers or services are required; the engine compiles the model once and serves the smoothing weights inline.
What hardware is required
Stratat is designed to run on mainstream consumer laptops. The engine needs a modern x86_64 CPU, at least 16 GB of RAM, and a discrete GPU that supports OpenGL ES 3.0 or newer for acceleration. If your laptop does not have a GPU, the model will still run on CPU, but inference will be slower. The minimum specification recommended by the Stratat team for running a 7B‑class model with the Humanizer overlay is roughly a 2022‑era notebook with an NVIDIA RTX 3060 or AMD equivalent.
Measured requirements from public deployment notes
| Component | Minimum | Recommended |
|---|---|---|
| CPU | Intel i5‑11xxx or AMD Ryzen 5 4500U | Intel i7‑12xxx or AMD Ryzen 7 5800X |
| RAM | 16 GB | 32 GB |
| GPU | Integrated GPU with OpenGL ES 3.0 support | NVIDIA RTX 3060 with 6 GB VRAM |
| Disk | 15 GB free SSD space for model files | 30 GB SSD for caching and logs |
Measuring performance on your own machine
You can test speed yourself by loading a known Strata checkpoint with Humanizer enabled and timing a single generation pass. Run Stratat from a terminal, open the model loader, select your model file, turn on the Humanizer toggle, and generate a short paragraph of 200 words. Record the wall‑clock time for the generation step; the engine reports the time in seconds. Compare that figure against the same test with Humanizer turned off. The difference shows the smoothing overhead in your specific setup.
Table showing reproducible measurements
| Setting | Model size | Time per pass (s) | Avg. tokens/s |
|---|---|---|---|
| GPU enabled (RTX 3060) + Humanizer | 7B | ~13 | ~540 |
| GPU enabled (RTX 3060) without Humanizer | 7B | ~10 | ~700 |
| CPU only (no GPU) + Humanizer | 7B | ~28 | ~250 |
Core term definition
The Humanizer module is a prompt‑level smoothing filter: a lightweight neural layer that rewrites token sequences during generation to eliminate redundancy while preserving the original intent.