Downloads

Model tools

Check a model before you trust it: where it gives way under pressure, what its weights look like, which concepts it holds ready to say. Every tool runs on your own computer, and nothing leaves it.

2026-10-06

Pivot Benchmark 0.9.0preview

You are about to put a local model in front of real people. Do you know what it takes to talk it into something it shouldn't do? Pivot Benchmark tells you before your users find out. A second model sits down with yours, starts with flattery, and then turns the conversation one word at a time: praise curdles into insult, a harmless remark hardens into an accusation, an honest request turns crooked. Every ladder runs beside a neutral control, so you get a measurement, not an anecdote. You see the exact dose that broke your model, how it broke, whether it recovered, and whether it gives way faster once it has given way before. Compare models, compare versions, and pick the one that holds. Free to download, no account, no licence key, and no installer on Windows.

You need: a local model served by llama.cpp, and your own Gemini API key.

About Pivot Benchmark · How to install

Windows 0.9.0-preview
Windows 10/11, 64-bit (.zip) · 191.4 MB
Unzip anywhere and run PivotBenchmark.exe. No installer; your settings and results go to %APPDATA%\Pivot Benchmark.
SHA-256d36b2647e7d74606808380a8056620320dd31ce172d0063f4d713f9c124c02c8
Download
Ubuntu 0.9.0-preview
Ubuntu 22.04 / 24.04, amd64 (.deb) · 164.7 MB
sudo apt install ./pivot-benchmark_0.9.0-preview_amd64.deb, then start it from the applications menu or with pivot-benchmark.
SHA-256e362d198ff2bf1960c99ecbd3c09eaff1daa6ae1155bb4de4cde9af8b9250529
Download

Before you run it

These builds are not code-signed. On Windows, SmartScreen shows "Windows protected your PC": choose More info, then Run anyway. Verify the file first and compare the result with SHA256SUMS.txt from the same release.

# Windows (PowerShell)
Get-FileHash .\PivotBenchmark-0.9.0-preview-windows-x64.zip -Algorithm SHA256

# Ubuntu
sha256sum pivot-benchmark_0.9.0-preview_amd64.deb

2026-10-06

PSE Analyzer 0.3.1

You downloaded a Q4 to save memory. What did it cost you? PSE Analyzer shows you, matrix by matrix. Every weight matrix has a spectrum, and the spectrum tells you where training left its mark, where the weights are still close to noise, and what went missing in compression. Put two precisions of the same model side by side and see which layers paid the price before you ship the smaller one. It reads a GGUF file one matrix at a time and never loads the whole model, so an ordinary CPU and a laptop's RAM are enough. Add torch and it goes deeper: the model's thinking layer by layer, the dictionary of concepts in its activations, and the handful it holds in mind. Every result is sealed in a dossier and never leaves your computer. Free to download, for Windows and Ubuntu.

You need: a GGUF model file. torch only for lens, sae, jspace and ppl.

About PSE Analyzer · How to install

Windows 0.3.1
Windows 10/11, 64-bit, single file (.exe) · 58.2 MB
One self-contained file. Run it from a terminal, or double-click to open the web view. It unpacks to a temporary folder on every start, about twelve seconds.
SHA-2561d47c403b3219f70ae36b0b000e8158a59bbfb2efe2cd9cdb43a68ba38068527
Download
Windows 0.3.1
Windows 10/11, 64-bit, folder build (.zip) · 59.5 MB
Unzip and run PSE-A.exe inside the folder. Starts in half a second; keep the whole folder together.
SHA-25637f64f3d9224de86ed3ae3951b32da36d737241d02bb0251504b6f1bfe52e5c6
Download
Ubuntu 0.3.1
Ubuntu 22.04 / 24.04 (.deb, no compiled code) · 0.3 MB
sudo apt install ./pse-a_0.3.1_all.deb, then pse-a doctor. The first run builds a Python environment and needs the internet once.
SHA-25694577c6c2100dc4616d6665ca7854490f4317f31a1e2fe5ac7114ea81cce36c8
Download

Before you run it

These builds are not code-signed. On Windows, SmartScreen shows "Windows protected your PC": choose More info, then Run anyway. Verify the file first and compare the result with SHA256SUMS.txt from the same release.

# Windows (PowerShell)
Get-FileHash .\PSE-Analyzer-0.3.1-windows-x64.exe -Algorithm SHA256

# Ubuntu
sha256sum pse-a_0.3.1_all.deb

2026-10-12

PSE Cartographer 0.2.0 beta

Maps, layer by layer, the concepts a Qwen, Llama or Mistral model holds ready to say. Open a local GGUF, feed it prompts and watch those concepts shift, and see the layer where the model suddenly commits. Around that sits the full PSE Analyzer toolset: the spectrum of every weight matrix, what quantization changed, what the model would say at each layer, and a dictionary of the concepts packed into its activations. Every result is sealed with a hash and told twice, in numbers and in plain words. A model of your choice writes the findings and conclusions, and every claim must cite the sealed result it comes from: Claude, Gemini, Mistral, Groq, or a local GGUF, fully offline. It runs on an ordinary CPU, with no GPU and no PyTorch; a 1B model maps in about ten minutes. Free and open source (GPL-3.0), for Ubuntu and Windows. Release notes

Ubuntu / Debian (.deb)
pse-cartographer_0.2.0_amd64.deb · 112.4 MB
SHA-25643258b508cd1ebe9b46a1eabc18c413ace1739c59926fa22aece0a87da9485a6
Download
Windows (.exe)
PSECartographer-Setup-0.2.0.exe · 79.4 MB
SHA-2569edad20ba81cb8aab14b8d00c2ddbd1fa59f4e7fdfbe06cf66b43df67d65faef
Download

2026-10-12

PSE Cartographer Max 0.4.0 beta

The same map for eight model families, plus a second one for mixture-of-experts models: which expert the router picks for each prompt. Open a local GGUF from Qwen, Llama, Mistral, Gemma, Phi, StableLM, OLMo or Granite and see, layer by layer, which concepts the model holds ready to say, how they shift with each prompt, and the layer where it suddenly commits. For Granite MoE and OLMoE, watch how the router's choice moves and which concepts travel with which expert. Around it sits the full PSE Analyzer toolset: the spectrum of every weight matrix, what quantization changed, what the model would say at each layer, and a dictionary of the concepts packed into its activations. Every result is sealed with a hash and told twice, in numbers and in plain words. A model of your choice writes the findings and conclusions, and every claim must cite the sealed result it comes from: Claude, Gemini, Mistral, Groq, or a local GGUF, fully offline. It runs on an ordinary CPU, with no GPU and no PyTorch; a 1B model maps in about ten minutes. Free and open source (GPL-3.0), for Ubuntu and Windows. Release notes

Ubuntu / Debian (.deb)
pse-cartographer_0.4.0_amd64.deb · 114.8 MB
SHA-256d1abbe2cb059b34f797ef059f26a46111fa82e62e2ab783190971ef1cb9ea36f
Download
Windows (.exe)
PSECartographer-Setup-0.4.0.exe · 81.4 MB
SHA-25674067437003e204825088b5ec92174f32aaa3d08b4b1c5fbd1f3d00c42725aad
Download

2026-10-12

PSE J-Lab 0.1.0 beta

A neural network is a function; its Jacobian is where all its sensitivity lives. PSE J-Lab opens that up on small nets you can actually watch. Five labs compute the input and parameter Jacobian exactly, then check every number against finite differences and, if you have torch, against autodiff too, so a result is trustworthy by construction. It ships with MNIST and FashionMNIST, seals every run with a hash and a plain-language report, and keeps two Guardians on the process. torch is an optional oracle, never required. Free and open source (GPL-3.0), for Ubuntu and Windows. Release notes

Ubuntu / Debian (.deb)
pse-jlab_0.1.0_amd64.deb · 119.8 MB
SHA-25679871408b2f090f3832654dde94c5e86fe43de57f6abfbfa96eebab75cab77e6
Download
Windows (.exe)
PSEJLab-Setup-0.1.0.exe · 92.1 MB
SHA-256f74b5787c5ba49a30988d34aed96fabbe6cf7be030270ae7c92a643b8a4ea130
Download

Before you run it

Check the file first

These builds are not code-signed yet. On Windows, SmartScreen shows "Windows protected your PC": choose More info, then Run anyway. Compare the SHA-256 above with sha256sum <file> on Ubuntu or Get-FileHash .\<file> -Algorithm SHA256 in PowerShell. All files come from github.com/zeravatechnologies.