Home Corpus Catalog Custom Training About & FAQ
Home About & FAQ

About Anubis AI

What is Anubis AI?

Anubis AI is a deterministic, auditable AI inference engine built entirely from scratch in C++17 and CUDA. No PyTorch. No TensorFlow. No external ML dependencies. Every operation is reproducible, traceable, and hash-audited.

The engine is designed to be lean at the base and deep at the edge. The base model knows how to reason, explain, and communicate in multiple languages. Domain knowledge is delivered through small, swappable adapter packs — rather than baking all knowledge into a single massive model.

Independently built and independently funded by a solo developer. Every line of C++ and every CUDA kernel, written by hand.

Company

Legal entity: Control AVS Technologies LLC
State: Texas
Product: Anubis AI
Main site: anubisai.us

How adapters work

1

Base model — reasoning engine

The Anubis base model (~832M–1.6B parameters) is trained to reason, explain, follow instructions, and communicate across 10+ languages. It contains no domain knowledge by design.

2

Adapter — domain knowledge layer

A small (5–50M parameter) adapter is trained on curated domain content. The adapter teaches the base model what to reason about — without modifying the base model itself.

3

Runtime — on-device inference

Load the base model + adapter into the Anubis runtime. Query in any language. Swap adapters to change domains. No cloud, no internet dependency after download.

4

Deterministic & auditable

Every inference is reproducible. Every operation is hash-audited. Results are cryptographically traceable — no probabilistic drift between runs.

Frequently asked questions

A corpus pack is a professionally curated, tokenized knowledge base for a specific subject domain. It includes all source text, processed tokens, vocabulary mappings, and metadata needed to train an Anubis adapter or use the corpus directly in retrieval-augmented generation (RAG) pipelines. Think of it as a high-quality training dataset, cleaned and ready to use.
Available Now means a pre-trained adapter file already exists for that subject. You receive the trained adapter immediately after payment — no waiting.

Pre-Order means the corpus exists but the adapter has not been trained yet. Your purchase triggers adapter training in our pipeline. Delivery is typically 24–72 hours, depending on queue. You receive the same adapter quality — it's just trained on demand rather than in advance.
Adapter files (.anubis format) require the Anubis inference engine to run. The base Anubis runtime will be available for download separately.

Raw corpus packs (tokenized text, vocabulary data) can be used with any compatible training pipeline, though they are optimized for the Anubis tokenizer and model format.

If you're evaluating for enterprise deployment, contact us for technical requirements and compatibility details.
The Anubis engine supports multiple hardware tiers:
  • Apple silicon (M1+): iPad Pro, MacBook — primary platform. ~1 GB memory for Q8 model.
  • Windows with NVIDIA GPU: 16 GB RAM minimum. ~1.85 GB VRAM for FP16 inference.
  • Windows CPU (AVX2): No GPU required. ~1 GB for Q8 model.
  • Enterprise server: CUDA-capable GPU with 16+ GB VRAM for larger model variants.
All corpus packs are built from authoritative, quality-filtered sources:
  • US Government publications (NIH, CDC, NIST, FAA, IRS, HHS, CFPB) — public domain
  • Project Gutenberg — public domain literature and reference works
  • Open Educational Resources (OER) — CC-BY licensed textbooks
  • OpenStax textbooks — CC-BY licensed open textbooks
  • Published academic standards and specifications

Each pack includes a quality report with source citations and corpus statistics.
The corpus packs are tokenized for the Anubis tokenizer (Etymo morpheme vocabulary) which is specific to the Anubis training pipeline. Raw source text is derived from public domain and open-licensed sources — you can access those independently. If you need raw text exports for use with other systems, contact us to discuss licensing options.
You own the trained adapter. We train on your corpus, deliver the adapter file, and delete your source corpus after delivery. We do not retain, redistribute, or use your proprietary content for any other purpose. See our Terms of Service for the full data handling policy.
Adapters are delivered as .anubis binary checkpoint files with SHA256 block verification. Each delivery includes a manifest.json with metadata, checksums, and the training configuration.
For pre-trained adapters (Available Now): refunds available within 24 hours of purchase if the file is defective or cannot be downloaded.

For pre-orders and custom training: refunds available before training begins. Once training is complete and the adapter has been delivered, we cannot issue refunds for digital goods.

See the full Refund Policy for details.
Payments are processed securely through Stripe. We do not store or handle your payment card information directly. Stripe accepts major credit cards and debit cards. For enterprise orders or custom payment arrangements, contact us.

Still have questions?

Reach out directly — we respond within 24 hours.

Email contact@anubisai.us