Frequently Asked Questions.

Yes. The core engine is built on llama.cpp with custom bindings that have ZERO network capability enabled by default. The software runs entirely in user-space RAM. You can verify this by running the application in an air-gapped machine or by inspecting our open-source bootloader code.

Our "Zero-Trace" architecture minimizes writes to the host disk. It operates as a portable application. However, Windows OS may create temporary pre-fetch files or registry entries for any connected USB device. We provide a cleanup script for advanced users to scrub these OS-level artifacts.

WARNING: AI models can hallucinate. SigmaNex is a decision support tool, not a doctor.

To mitigate this, our "Medical" preset uses Retrieval Augmented Generation (RAG) tied to a static, verified embedding database of TCCC and Red Cross guidelines. The AI is instructed to cite these static documents rather than generating creative text. Always cross-reference if possible.

Minimum: 8GB RAM, Quad-Core CPU (AVX2 support).
Recommended: 16GB RAM, NVIDIA GPU (6GB+ VRAM).
The system dynamically offloads layers to the GPU if available, or falls back to optimized CPU inference.

The Alpha release ships with fine-tuned variants of Llama-3 (8B) and Mistral (7B). These are quantized to Q4_K_M precision to balance perplexity and memory bandwidth. Users with >16GB VRAM can hot-swap these for larger parameter models (e.g., Command R+ or Llama-3 70B) by placing GGUF files in the /models directory.

Yes. SigmaNex is model-agnostic. As long as the model is compatible with the llama.cpp backend (GGUF format), you can drop it onto the USB drive. Our bootloader will automatically detect new weights and offer them in the "Presets" menu.

SigmaNex is a tool for exercising Digital Sovereignty. Running inference locally on your own hardware is generally a protected activity in most jurisdictions. However, you differ from SaaS users in that you are the data controller. We advise checking local laws regarding encryption and dual-use software.

Impossible. We have zero visibility into your runtime environment. There is no telemetry module in the kernel. No "crash reporting" pinging our servers. If the software crashes, it crashes silently to protect your operational security.

We are currently self-funded by the founding team. We accept donations to cover the cost of cloud GPUs used for training the finalized models. We do not sell user data, we do not run ads, and we do not have VC backers demanding growth metrics.